OpenMAMA Developer's Guide

Introduction and Architecture

The MAMA (Middleware Agnostic Messaging) API is a subscription based messaging API with publish/subscribe semantics, which provides a lightweight abstraction on top of a variety of underlying messaging middlewares. The OpenMAMA API provides developers with a common interface to the underlying messaging API, allowing migration from one messaging API to another without any code changes to applications built using the API.

The API provides an asynchronous, event-driven programming model. You use the API to provide callbacks where required. Data is propagated back to the registered application via these callbacks in response to dispatching of events from event queues.

Market data semantics are added through the use of the OpenMAMDA API. The OpenMAMA API also provides extra features when used on the NYSE TechnologiesMarket Data Infrastructure, such as:

• Subscription throttling
• Entitlement enforcement
• Initial values/recaps
• Data quality
• Refresh messages
• Group subscriptions

The document details the major objects within the API and their most common usage.

Operating Systems

The following operating systems are currently supported:

• Linux
• Windows

Middlewares

The following middlewares are currently supported by the OpenMAMA project, but many more are supported via third parties:

• Qpid

API Language Implementations

The various language implementations for the OpenMAMA API expose the same top level objects with broadly the same programming interface. Unless otherwise specified, allocation of OpenMAMA objects and creation of OpenMAMA objects are two separate steps. The first step allocates the memory and the second initializes the object. Each of the API implementations is thread safe and thread aware. All functions/methods in the API, across all language implementations and transports, exhibit the same behavior, unless otherwise stated.

The following language implementations are currently supported:

• C
• C++
• C#
• JNI

The C API naming convention is used to identify objects and the functions that operate on those objects. Each object has its own header file containing definitions for all operations supported by that object. The naming convention is as follows:

• Objects: mama<Type>
• Functions: mama<Type>_operation

All structures are defined as opaque types. As such they cannot be allocated directly by users of the API. All memory management for a OpenMAMA object in C is controlled via calls to the appropriate functions, for example:

• mama<Type>_allocate()
• mama<Type>_destroy()
• mama<Type>_deallocate()

All functions in the C API return a mama_status value. A value of MAMA_STATUS_OK indicates successful invocation. Any other result is an error condition. The mama_status enum is defined in mama/status.h.

mama_open() must be called before any other object creation. This call is responsible for initializing the API internally.

Sample applications written using the API are located in the examples directory provided with the distribution. Pre-compiled versions of these example programs are located in the bin directory. Example programs are detailed in the Example Programs section. For clarity, the sample snippets of code do not check the return values from function/method calls. In a production application it is recommended that all return codes are checked for success before proceeding.

Use of the API is not restricted to the NYSE Technologies Market Data Platform. We provide the concept of ‘Basic’ subscribers and publishers which allow users of the API to publish and subscribe to topics outside of the Market Data Platform (details provided in relevant sections).

All callback functions/methods provide access to closure data in order to allow applications to associate contextual information in their own applications with callbacks invoked from within the OpenMAMA API.

Using the API

The following is a high level overview of the steps required to implement the most common use of the OpenMAMA API: to write a market data subscribing application. Each step and their associated features are detailed in the following chapters.

1. Load the bridge(s). At least one bridge object must be created before OpenMAMA is opened. Bridges can only be loaded at this time.

2. Initialize the API. Before creating any OpenMAMA objects it is necessary to initialize the API. Open OpenMAMA by calling the mama_open() function. This call sets up a number of internal processes within the API that are required for successful use of the API. Note The call to mama_open() is reference counted and each call must have a corresponding call to mama_close().

3. Create event queue(s) (optional). Users of the API can use the default internal event queue for a bridge. If multi-threading is required, or more control over the dispatching of events, separate event queues can be created using MamaQueue(). See the section on Events and Queues for details. Note It is not necessary for an application to create its own event queues.

4. Create transport(s). An application requires that at least one transport object has been created. The mamaTransport object is used to define the properties for the underlying middleware’s communication protocol. There must be an entry in the mama.properties file for each transport created. See the section Transports for details.

5. Fetch data dictionary. A data dictionary is typically required in order to obtain complete information (name, fid and data type) regarding fields within messages. See the section on the OpenMAMA Dictionary for details.

6. Create subscriptions. Create a subscription object, mamaSubscription, for each symbol known to the application at startup. See the section on Subscriptions for details.

7. Start dispatching. Once all subscriptions have been created, start dispatching on the default event queue for a bridge, and any other queues that have been created. You can continue to create further intra-day subscriptions.

8. Shut down. Objects must be destroyed in the following order when shutting down the application:

   1. Stop dispatching on any event queues (also call Mama::stop())
   2. Destroy event object (Subscriptions, timers, io objects)
   3. Destroy all event queues
   4. Destroy all transports
   5. Call mama_close()

Event objects can be created and destroyed at runtime, however, queues must not be destroyed before all event objects that use those queues are destroyed. It is recommended that OpenMAMA objects are destroyed in this order, as it will result in the same behavior on all middlewares currently supported.

Note This applies only to shutdown. Event objects can be destroyed at any point during the life of an application.

Object Summary

Here is a list of the main OpenMAMA Objects that are available for use within an OpenMAMA based application:

Installation

Linux

Installation

OpenMAMA for Linux may be distributed via a tarball, or as a .deb / .rpm package, but the contents and structure remain the same when extracted:

Before running OpenMAMA:

1. Set the WOMBAT_PATH variable to include the path to the directory containing mama.properties:

   $ export WOMBAT_PATH=/opt/openmama/config
   

2. Set the LD_LIBRARY_PATH variable to include lib or alternatively add an entry to /etc/ld.so.conf:

   $ export LD_LIBRARY_PATH=/opt/openmama/lib:$LD_LIBRARY_PATH 
   

Running Example Programs

When you have extracted the API and set the environment variables as described above, you can run the example programs. For the example programs to receive any data there must be a publisher somewhere on the network.

To run the mamalistenc program, change to the bin directory and run the following command:

$ ./mamalistenc -S SOURCENAME -tport TRANSPORT -s SYMBOL 

Where:

• SOURCENAME is the source name
• TRANSPORT is the transport name
• SYMBOL is the symbol you wish to subscribe to

The transport should be defined in both mama.properties and the publisher configuration. The source should be defined in the publisher configuration.

The other example programs use similar options to mamalistenc. For a full list of options for a particular program, use the -? option. For example, to see all the options for bookviewer, run the following command:

$ bookviewer -?

Compiling Example Programs

The examples are found in examples/mama/language. They can be compiled using make and Makefile. sample. The variable API_HOME has to be set to the directory where the API was installed. For example, to build mamalistenc, run the following command:

$ make -f Makefile.sample API_HOME=/opt/openmama mamalistenc 

Where:

• API_HOME is the full path to the directory where the API was installed

To build all the example programs, run make without a target:

$ make -f Makefile.sample API_HOME=/opt/openmama

See the files in that directory for a list of supported example applications for your chosen language.

Windows

Installation Steps

OpenMAMA for Windows is distributed as a zip file, but the contents and structure remain the same when extracted:

Before running OpenMAMA:

1. Set the WOMBAT_PATH variable to include the path to the directory containing mama.properties:

   $ export WOMBAT_PATH=c:\mama\config
   

2. Set API_HOME to the directory where you extracted MAMA:

   $ set API_HOME=c:\mama 
   

3. Set the PATH variable to include bin directory.

   $ set PATH=c:\mama\bin;%PATH%
   

This list must also include any proprietary third party libraries and their dependency .dlls too, if appropriate.

Bridges

Middleware Bridges

OpenMAMA supports the different middlewares through the use of bridge objects. Multiple bridges can be loaded at any one time, one for each middleware, meaning that a single OpenMAMA application can support more than one middleware concurrently.

Bridge objects must be created at startup, before MAMA’s open function is called. Once initialized, bridge objects are passed as parameters when creating transport, queue and queue group objects. Any further objects (such as subscriptions or timers) that use a transport or queue automatically use the same middleware and bridge. MAMA’s start and stop functions also take a bridge as a parameter to start dispatching on the default event queue for that bridge.

The bridge implementation libraries shipped with OpenMAMA contain all the middleware specific functions. These bridge libraries can either be linked into the application at link time, or dynamically loaded at runtime (all languages), as detailed in the following sections.

Using Linked Bridge Libraries

If the bridge implementation libraries are linked into the OpenMAMA application, either statically or dynamically, then a bridge can be loaded using one of the following methods:

• Using one of the middleware specific load functions.
• Through a macro that takes the middleware identifier as a parameter.

An application will fail to compile if it tries to use a middleware bridge that has not been linked in.

Loading Bridge Libraries at Runtime

Bridge libraries can be dynamically loaded at runtime. This is different from dynamic linking as the bridge libraries are not actually linked into the application. The bridge to be used is decided at runtime. Using dynamic loading offers the greatest flexibility as it means that applications do not have to be recompiled to use different middlewares. This is the method that the example applications use.

If the application tries to load a bridge library that is not available, or if the middleware libraries for that bridge are not available, then the load will fail. A path to the bridge libraries must be available from LD_LIBRARY_PATH, for Unix systems, or PATH, for Windows systems.

To load bridges in:

mamaBridge bridge = NULL; 
bridge = Mama::loadBridge ("qpid"); 

mamaBridge bridge = NULL; 
bridge = mama_loadBridge (&bridge, "qpid"); 

Mama.loadBridge("qpid");

Mama.loadBridge("qpid");

Loading Bridge Libraries at Runtime from a Specified Location

OpenMAMA also allows the user to load the bridge libraries from a specified path. If no path is cited, loadBridge functionality defaults to using the system’s library search paths. The specified path must use the appropriate path separator for the OS i.e. “/” for Unix systems, “\” for Windows systems.

To load in bridge libraries at runtime with path:

mamaBridge bridge = NULL;
bridge = Mama::loadBridge ("qpid", "/opt/openmama/lib");

mamaBridge bridge = NULL;
mama_loadBridgeWithPath (&bridge, "qpid", "/opt/openmama/lib");

MamaBridge bridge = null;
bridge = Mama.loadBridge("qpid", "/opt/openmama/lib");

MamaBridge bridge = null;
bridge = Mama.loadBridge("qpid", @"C:\openmama\lib");

Payload Bridges

OpenMAMA supports different payloads through the use of bridge objects. Multiple bridges can be loaded at any one time, one for each payload, meaning that a single OpenMAMA application can support more than one payload concurrently.

Payload bridge objects follow the same rules for loading as middleware bridges.

Loading Payload Bridges at Runtime

Payload bridge libraries can be dynamically loaded at runtime in the same way as a middleware bridge.

To load payload bridge libraries at runtime:

mamaPayloadBridge payload = NULL;
payload = Mama::loadPayloadBridge ("qpidmsg");

mamaPayloadBridge payload = NULL;
mama_loadPayloadBridge (&payload, "qpidmsg");

MamaPayloadBridge payload = null;
payload = Mama.loadPayloadBridge("qpidmsg");

MamaPayloadBridge payload = null;
payload = Mama.loadPayloadBridge("qpidmsg");

Default Payloads

A middleware bridge may specify which payload to use, which is identified during the middleware bridge loading process. When specified, OpenMAMA tries to load the requested payload bridge. If the payload is not available the middleware bridge continues to load, and the payload load failure is logged.

The first payload bridge that is successfully loaded is marked as the default payload. This can be overridden programmatically.

To set the default payload:

// Not currently supported in the C++ API

mama_setDefaultPayloadBridge (‘Q’); 

// Not currently supported in the Java API

// Not currently supported in the C# API

Using Payload

The default payload is used when no payload bridge is explicitly stated. There are three options available for specifying payload creation:

Using the default

MamaMsg msg = new MamaMsg();
msg->create();

mamaMsg* msg = NULL; 
mamaMsg_create(&msg); 

MamaMsg msg = new MamaMsg();

MamaMsg msg = new MamaMsg();

Using the payload ID

MamaMsg msg = new MamaMsg();
msg->createForPayload(‘Q’); 

mamaMsg* msg = NULL; 
mamaMsg_createForPayload(&msg, ‘Q’); 

MamaMsg msg = new MamaMsg('Q');

MamaMsg msg = new MamaMsg('Q');

Using the bridge structure

MamaMsg msg = new MamaMsg();
msg->createForPayloadBridge(payloadBridge); 

mamaMsg* msg = NULL; 
mamaMsg_createForPayloadBridge(&msg, payloadBridge); 

MamaMsg msg = new MamaMsg(payloadBridge);

MamaMsg msg = new MamaMsg(payloadBridge);

Properties

There are a number of mechanisms in place by which properties for the OpenMAMA API, and the underlying messaging middleware specific APIs, can be specified.

The default behavior for the API (for all languages) is to look for a file named mama.properties in the directory specified by the WOMBAT_PATH environment variable. If present, the specified properties file is loaded on application startup when MAMA’s open function is called. OpenMAMA accesses the properties file and looks for all the possible properties and loads them in. Misspellings and omissions are not highlighted on startup. Please look at the documentation for the messaging middleware being used for more detail on, and an explanation of, transport level properties.

Alternatively, users of the API can override this behavior and can specify a file name and location using MAMA’s open with properties function. The fully qualified path to the directory is required.

Setting Properties at Runtime

Properties can also be specified at runtime. This approach can be used to override existing properties specified in mama.properties, or to add new properties to the API. All properties should be specified prior to creating any transport objects.

Note: Calling mama_open or mama_openWithProperties overrides any existing same-named properties set prior to these calls.

Properties are typically interpreted when objects such as transports or subscriptions are created, therefore changing these at runtime has little or no effect.

The following example illustrates the setting of the router location for the Avismessaging middleware. Properties are always specified as strings in all API versions.

Setting properties at runtime

Mama::setProperty("mama.payload.default", "qpidmsg");

mama_setProperty ("mama.payload.default", "qpidmsg");

Mama.setProperty("mama.payload.default", "qpidmsg");

Mama.setProperty("mama.payload.default", "qpidmsg");

Regardless of the mechanism used to specify the properties, or the specific language implementation used, the properties must always have the same format. An example of a mama.properties file is provided in the examples directory as part of the release structure, which highlights and describes the most commonly specified properties for each supported middleware.

Transports

The underpinning object in any OpenMAMA application is the MamaTransport object. The MamaTransport object defines the network protocol-level parameters over which OpenMAMA distributes data. Transports effectively provide scope for data identifying the underlying protocols and their values for data delivery. This object is an isolated communication channel for data traffic. Transports specify the communication channels to use for making subscriptions and publishing data through the API.

A single OpenMAMA application can define multiple transport objects if more than one set of physical transport channels exist on the network.

When a transport object is created it is given a name identifier, which is used internally to obtain any configuration parameters that have been specified on application startup. These properties can be specified in a variety of ways, the most common being in mama.properties. If the name is not found or null, then the defaults specified by OpenMAMA for that middleware are used. The defaults can be overridden by creating transport entries in the properties file with the name default.

Creating a Transport

The following example code illustrates how to create a transport object using the identifier “avis_tport” to locate relevant properties from the list specified to the API upon initialization.

MamaTransport* transport = new MamaTransport(); 
// Set runtime properties before "creating" transport here ...
transport->create ("transport_name", bridge);

mamaTransport transport;
mamaTransport_allocate (&transport);
// Set runtime properties before "creating" transport here ...
mamaTransport_create (transport, "transport_name", bridge);

MamaTransport transport = new MamaTransport();
// Set runtime properties before "creating" transport here ...
transport.create ("transport_name", bridge);

MamaTransport transport = new MamaTransport();
// Set runtime properties before "creating" transport here ...
transport.create ("transport_name", bridge);

Setting Transport Properties

Properties are typically set on a per transport basis within the API. Transport properties always follow this naming convention:

mama.<middleware>.transport.<transport name>.<property name>=<value>

where:

• <middleware> is the middleware bridge name
• <transport name> is the string identifier used when creating a transport object
• <property name> is the name of the property ad the messaging middleware level

The property name and property value can be separated by either a space (' ') or an equals sign ('=').

As an example for qpid:

mama.qpid.transport.pub.incoming_url=amqp://~127.0.0.1:7777

Transport Runtime Attributes

There are several transport runtime attributes that can be set after allocation.

The API provides throttling of subscription creation and the rate at which recaps can be sent from the API. The default values are 500 per second for initial values (subscription creation) and 250 per second for recap requests. Subscription requests are throttled on the default throttle and recaps on the recap throttle.

The throttles are created and can be configured on a per transport basis. The values for the throttle rates can be changed at runtime via the transport’s outbound throttle functions. A callback can also be supplied to register for any transport level events that occur. This is dependent on the underlying messaging middleware being used, as not all provide this type of information. E.g:

/* Change the value of the default throttle to 1000 msg/sec */ 
transport->setOutboundThrottle (1000.0, MAMA_THROTTLE_DEFAULT); 

/* Change the value of the recap throttle to 500 msg/sec */ 
transport->setOutboundThrottle (500.0, MAMA_THROTTLE_RECAP);

/* Set an optional transport level callback */ 
transport->setTransportCallback (callback);

/* Change the value of the default throttle to 1000 msg/sec */ 
mamaTransport_setOutboundThrottle (transport, MAMA_THROTTLE_DEFAULT, 1000.0); 

/* Change the value of the recap throttle to 500 msg/sec */ 
mamaTransport_setOutboundThrottle (transport, MAMA_THROTTLE_RECAP, 500.0);

/* Set an optional transport level callback */ 
mamaTransport_setTransportCallback (transport, callback);

/* Change the value of the default throttle to 1000 msg/sec */ 
transport.setOutboundThrottle (MamaThrottleInstance.DEFAULT_THROTTLE, 1000.0); 

/* Change the value of the recap throttle to 500 msg/sec */ 
transport.setOutboundThrottle (MamaThrottleInstance.RECAP_THROTTLE, 500.0);

/* Set an optional transport level callback */ 
transport.addTransportListener(new MamaTransportListener());

/* Change the value of the default throttle to 1000 msg/sec */ 
transport.setOutboundThrottle (MamaTransport.MamaThrottleInstance.MAMA_THROTTLE_DEFAULT, 1000.0); 

/* Change the value of the recap throttle to 500 msg/sec */ 
transport.setOutboundThrottle (MamaTransport.MamaThrottleInstance.MAMA_THROTTLE_RECAP, 500.0);

/* Set an optional transport level callback */ 
transport.setTransportCallback (new ListenerTransportCallback());

Load Balancing Transports

Note: This functionality is not available on all middleware bridges - please refer to third party vendors for support details.

A transport represents an isolated communication channel for data traffic. However, it is also possible to create a load-balanced transport, which makes available a set of channels. Different channels may be selected to balance the amount of traffic transmitted across this set. The selection could be made at random, or in a round-robin fashion, or in some other way according to user requirements.

Two types of load balancing scheme are possible:

• Client Load Balancing: In this case, a client selects one of the channels for use. Different clients may then use different channels.
• Subscription Load Balancing: In this case, a client may select one of the channels for each symbol it is registering an interest in. Different symbols may then use different channels.

Properties for Load Balancing (Vendor Specific)

Note that API calls are only available for the C API.

A number of properties may be used for load balancing:

lb<n>, lb_scheme, lb_shared_object

There are a number of different load balancing schemes available. Each scheme requires a set of transports to be created under the same name. This set of transports is then available for making different selections. The set of transports is defined by adding “lb” terms after the name of the transport. These terms need to be numbered consecutively from 0, without gaps in the numbering, as illustrated in the following example.

mama.wmw.transport.sub.lb0.subscribe_address_0=cache0
mama.wmw.transport.sub.lb0.subscribe_port_0=1457
mama.wmw.transport.sub.lb1.subscribe_address_0=cache1
mama.wmw.transport.sub.lb1.subscribe_port_0=1457 

Once the transport set is defined, one of two load balancing schemes may be chosen: either select a transport from the set of transports and use this transport to create all subscriptions (client load balancing), or select a transport from the set of transports each time a subscription is created (subscription load balancing).

Client Load Balancing

With client load balancing, a transport is selected at random and used for all subscriptions. Create a callback, such as the following, to override this random selection and instead make the selection from a shared object (dynamic link library) or directly in code:

void mamaTransportLbInitialCB (int numTransports, int* transportIndex);

This must pass back an index (0 <= transportIndex < numTransports). In code, this callback is set with the following function:

mama_status 
mamaTransport_setLbInitialCallback (mamaTransport            transport,
                                    mamaTransportLbInitialCB callback);

Subscription Load Balancing

The default behavior for the subscription load balancing scheme is to begin from the first transport for the first subscription, then select transports in round-robin for each subsequent subscription created. This initialization and round-robin selection may be overridden by using the mamaTransportLbInitialCB callback described above to choose the initial transport, and by creating a mamaTransportLbCB callback to make the selection in a shared object, dynamic link library, or directly in code, as follows:

void mamaTransportLbCB (int         curTransportIndex, 
                        int         numTransports, 
                        const char* source, 
                        const char* symbol, 
                        int*        nextTransportIndex);

This must pass back an index (0 <= transportIndex < numTransports). The source and symbol name are passed to assist the decision process, for example, splitting the subscriptions alphabetically. In code, this callback is set with the function:

mama_status 
mamaTransport_setLbCallback (mamaTransport     transport,
                             mamaTransportLbCB callback);

Summary of Properties

The load balancing schemes described above are controlled by a set of four properties:

• round_robin: This will round robin the transport used for each subscription.

  mama.wmw.transport.sub.lb_scheme = round_robin 
  

• static: This will select a transport at random to use for all subscriptions.
• api: This will allow the use of mamaTransport_setLbInitialCallback and mamaTransport_setLbCallback to set callbacks to override default behaviour.
• library: This will allow the use of a shared object (dynamic link library) with loadBalanceInitial and loadBalance entry points to override default behaviour.

  mama.wmw.transport.sub.lb_scheme = library
  mama.wmw.transport.sub.lb_shared_object = alpha.so
  

Restrictions on Load Balancing

In the event that one of the transports within a load-balanced group is down, the subscriptions will not automatically fail over to another transport within the group.

Load balancing is currently only currently available with selected middlewares. Please refer to your bridge provider’s documentation / support for details.

Resource Pool (v6.3.2+)

The OpenMAMA resource pool is a convenience layer around OpenMAMA core which allows the caller to create subscriptions, subscribe to messages and build applications without having to get involved in the nitty gritty of OpenMAMA object lifecycles.

Primary goals are to:

• Provide a new more convenient interface for managing subscriptions, reducing any barriers to entry for new users
• Increase accessibility of the OpenMAMA API for new users
• Provide a convenient and well tuned interface “out of the box” where tuning options can be configured rather than coded
• Provide a small interface which would be easy to port to other languages

We should always bear in mind that part of the goal here isn’t just to make OpenMAMA accessible for new users; it’s also to help market data vendors distribute their data more easily and risk free to new prospects. Core approaches may include:

• Using existing OpenMAMA objects as the currency of the interfaces provided here (e.g. _getSubscription may return a mamaSubscription type).
• Abstracting common performance tunings away from code and into configuration (e.g. thread pinning, subscription distribution to queues etc).

The idea is to reduce an OpenMAMA subscribing application to something which looks more like this:

// TODO: Not yet implemented

// Use a new "resource pool" to manage lifecycles of all pool-created resources,
// allocating queues, defaults etc from configuration
mamaResourcePool resourcePool;
mamaResourcePool_create(&resourcePool, "pool_name");

// Set up the event handlers for subscription
mamaMsgCallbacks callbacks;
memset(&callbacks, 0, sizeof(callbacks));
callbacks.onMsg = subscriptionOnMsg;
callbacks.onCreate = subscriptionOnCreate;
callbacks.onError = subscriptionOnError;

// Actually... subscribe to something via something convenient like a URI (Additional API)
const void* closure = NULL;
mamaResourcePool_createSubscriptionFromUri(resourcePool,
                                           &subscription,
                                           "bridge://transport/source/topic",
                                           &callbacks,
                                           closure);

// Block and begin receiving callbacks
mama_start();

// TODO: Not yet implemented

// TODO: Not yet implemented

The idea is to use configuration such as below to manage things like the number of queues to use, how do load balance etc:

# The number of threads / queues to use in the pool’s queue group
mama.resource_pool.<pool name>.queues = 2

# Any preferred regex for assigning URIs to threads on creation
mama.resource_pool.<pool_name>.queue_0.regex = ^.*[/]b*[0-9A-M].*$
mama.resource_pool.<pool_name>.queue_1.regex = ^.*[/]b*[N-Z].*$

# Name of OpenMAMA queue thread to create in the queue group (to work
# alongside mama.thread_affinity (will be <thread_prefix>_N where N is
# the queue index).
mama.resource_pool.<pool_name>.thread_name_prefix = <pool_name>

# List of bridges to load (by default will try to load all available)
mama.resource_pool.<pool_name>.bridges = <bridge X> <bridge Y>

So in this case, the mamaResourcePool will:

• Create queues for subscriptions (which could be configurable based on the pool name).
• Load all middleware bridges found on the PATH / LD_LIBRARY_PATH etc.
• Call mama_open() (if not already open).
• Create and manage internal stores of all managed subscriptions, transports, sources etc.

Then when a resource is created inside:

• Find / create transport bridge.
• Find / create transport.
• Find / create source.
• Find / create subscription (e.g. multiple callbacks for same subscription).
• Find / create anything which is required to acquire the requested resource.
• Be convenient and usable, but still flexible (it would return existing OpenMAMA objects so they can be manipulated).
• Take the heavy lifting and thread safety pain away from the application developer.

For a complete reference to the convenience functions available, please refer to the OpenMAMA reference guide for details.

There are some mandatory conventions which need to be adhered to in order to fully take advantage of resource pools:

• Transport names should be unique, even cross-bridges. Sharing the same transport name across multiple bridges should be considered bad practice, and allowing mamaResourcePool to be hindered because of this particular use case was considered too restrictive and inflexible. So with that in mind, mamaResourcePool and assocated APIs mandate that the same transport name cannot be used for multiple middlewares.

• Source names should be unique, even cross transports. Allowing multiple sources to appear across multiple transports reduced flexibility in the API, so with that in mind it was agreed that the same source may not be used on multiple transports when using the MAMA resource pool.

This allows us to open up API calls that don’t require passing bridges or transports around everywhere since they can be managed by the mamaResourcePool instead, for example:

// TODO: Not yet implemented

// This one will use the transport, source and topic to subscribe... with the mama resource pool
// finding / creating the transports and sources requested
mamaResourcePool_createSubscriptionFromComponents (resourcePool,
                                                   &subscription,
                                                   "transportname",
                                                   "sourcename",
                                                   "topicname",
                                                   &callbacks,
                                                   closure);

// This one will imply using the mama resource pool's default transport
mamaResourcePool_createSubscriptionFromTopicWithSource (resourcePool,
                                                        &subscription,
                                                        "sourcename",
                                                        "topicname",
                                                        &callbacks,
                                                        closure);

// This one will imply using the mama resource pool's default transport, and that transport's
// default source (or the pool's default source if provided)
mamaResourcePool_createSubscriptionFromTopic (resourcePool,
                                              &subscription,
                                              "topicname",
                                              &callbacks,
                                              closure);

// TODO: Not yet implemented

// TODO: Not yet implemented

Which would then depend on configuration such as this to define the default source / transport for the resource pool:

# Configure a default transport for this resource pool
mama.resource_pool.<pool name>.default_transport = <transport>

# Default source may be provided by either the resource pool, or the transport itself
mama.resource_pool.<pool name>.default_source = <source_name>

# New source level configuration to define the transport and bridge
mama.source.<source_name>.transport = <transport>
mama.source.<source_name>.bridge = <bridge>

# Default source may also be specified at a transport level
mama.transport.bridge.<transport>.default_source = <source_name>

This means that application developers don’t need to go through the motions of URI construction to build everything they need for a transport and instead they can defer that boilerplate to configuration instead.

An added benefit of adopting this approach means that an API such as the following may safely be used to generate “pick lists” of available transports without fear of conflict or overlap since transport names are now configured-application-instance-unique:

// TODO: Not yet implemented

// Optionally gather all available transports defined in mama.properties for use in pick lists
char transportsFound[5][MAMA_MAX_TRANSPORT_LEN];
size_t found;
mama_getAvailableTransportNames (transportsFound, 5, &found);

// TODO: Not yet implemented

// TODO: Not yet implemented

Which would then be all you would need to use a mamaResourcePool to create and initialize that transport (no need to throw mamaBridge instances around):

// Gather just the transport for further metadata to be used in pick lists should need arise
mamaTransport transport;
mamaResourcePool_createTransportFromName (resourcePool,
                                          &transport
                                          "transport_name");

Or indeed simply let the mamaResourcePool just-in-time initialize that transport when the first subscription using it is instantiated.

Events and Queues

Events and queues are the core functionality of the OpenMAMA API, as they enable asynchronous, event-driven data processing. OpenMAMA applications register interest in events, from data arriving on a socket for a subscription or being informed of an elapsed timer, and execute application code through callbacks in response to them.

Each event created needs a corresponding callback function created that is invoked by the API once an event of the specific type occurs. A closure can also be specified for the event when it is being created. A closure is an arbitrary piece of data which is returned to the user in the callback. It provides a mechanism by which a user can associate context between the code in the callback and the application environment. Closures are specified with opaque data types.

Events, when they occur within the API, are placed onto an event queue. An event queue is the mechanism that controls the dispatching of events within the API. These events result in the invocation of a callback for that event type once the event has been dispatched from the event queue.

Dispatching an event involves removing the next available event from the event queue, identifying the event type, and invoking the corresponding callback registered for that event type.

Queues are represented by the MamaQueue object.

The API maintains a default internal data queue for each bridge which it uses for internal timers and controls such as the throttling of subscriptions. You can use the default event queue for a bridge when creating subscriptions, timers, and so forth. Dispatching on this queue starts once MAMA’s start function is called. In this case the API is essentially being used in single threaded mode. The default event queue should never be destroyed by the calling application. The call to MAMA’s start function is reference counted and each call must have a corresponding call to MAMA’s stop function. Dispatching stops when MAMA’s stop function is called for the final time. Both MAMA’s start and stop calls are thread safe.

When a queue is being actively dispatched, events on that queue may only be dispatched by the dispatching thread. If dispatching is stopped then an event object can be destroyed from any thread. This is true for both the default queue and user-created queues.

In OpenMAMA all data is propagated in response to events being dispatched from one or more event queues. Callbacks registered with the API are invoked on the threads dispatching on particular queues whenever an event is available. Internally, the middleware adds data onto the event queue and it is the responsibility of the application code to dispatch events from these queues in a timely fashion.

Accessing the Internal Event Queue

MamaQueue* queue = NULL;
queue = Mama::getDefaultEventQueue (bridge);

mamaQueue queue;
mama_getDefaultEventQueue (bridge, &queue)

MamaQueue queue = Mama.getDefaultQueue (bridge);

MamaQueue queue = Mama.getDefaultEventQueue (bridge);

Creating Queues

For multi-threaded dispatching and for more control over the de-queuing of events in the API you can create your own queues from which events can be manually dispatched, as illustrated in the following example.

MamaQueue* queue = new MamaQueue;
queue->create (bridge);

mamaQueue queue;
mamaQueue_create (&queue, bridge);

MamaQueue queue = MamaQueue (bridge);

MamaQueue queue = MamaQueue (bridge);

Destroying Queues

An OpenMAMA queue can only be destroyed if all the objects using it have been destroyed first (such as timers, inboxes and subscriptions).

Objects like these are destroyed asynchronously which means that there is a time delay between calling the ‘destroy’ function and the object actually being destroyed. The object is deemed to be destroyed whenever it is impossible for further events to be placed onto the queue on its behalf.

Each time one of these objects is created, a lock count is incremented on the associated queue, which is only decremented when that object is fully destroyed.

How to Destroy a Queue

An OpenMAMA queue can be destroyed using the queue’s destroy function shown in the following example. If there are any open objects on the queue then an error code is returned.

/* Attempt to destroy the queue. */
try {
    mama_status status = queue->destroy (); 
} catch (MamaStatus e) {
    printf ("Can't destroy queue: %s\n", e.toString ()); 
}

/* Attempt to destroy the queue. */ 
mama_status status = mamaQueue_destroy(queue); 
if(status == MAMA_STATUS_QUEUE_OPEN_OBJECTS) 
{ 
    printf ("Can't destroy queue as there are open objects.\n"); 
}

/* Attempt to destroy the queue. */ 
try {
    queue.destroy ();
} catch (MamaException e) {
    System.out.println (e);
}

/* Attempt to destroy the queue. */ 
try {
    queue.destroy ();
} catch (MamaException e) {
    Console.WriteLine ("Can't destroy queue: {0}\n", e.Message );
}

If the objects using the queue have had their ‘destroy’ functions called but have not been fully destroyed, then the destroyWait function can be used. This blocks and processes messages from the queue until all objects have been fully destroyed:

/* Block until all objects have been destroyed. */ 
try {
    mama_status status = queue->destroyWait (); 
} catch (MamaStatus e) {
    printf ("Can't destroy queue: %s\n", e.toString()); 
}

/* Block until all objects have been destroyed. */ 
mama_status status = mamaQueue_destroyWait (queue); 
if(status == MAMA_STATUS_QUEUE_OPEN_OBJECTS) 
{ 
    printf ("Can't destroy queue as there are open objects.\n"); 
}

/* Block until all objects have been destroyed. */ 
try {
    queue.destroyWait ();
} catch (MamaException e) {
    System.out.println (e);
}

/* Block until all objects have been destroyed. */ 
try {
    queue.destroyWait ();
} catch (MamaException e) {
    Console.WriteLine ("Can't destroy queue: {0}\n", e.Message );
}

The destroyTimedWait function behaves in the same manner as destroyWait but only blocks for the supplied timer period. Once this period elapses a timeout error is returned if all open objects have not been destroyed.

/* Block until all objects have been destroyed. */ 
try {
    mama_status status = queue->destroyTimedWait (6000); 
} catch (MamaStatus e) {
    printf ("Can't destroy queue: %s\n", e.toString()); 
}

/* Process messages for 6 seconds. */ 
mama_status status == mamaQueue_destroyTimedWait (queue, 6000); 
if(status == MAMA_STATUS_TIMEOUT) 
{ 
    printf ("Timed out waiting for queue to be destroyed.\n"); 
}

/* Block until all objects have been destroyed. */ 
try {
    queue.destroyTimedWait (6000);
} catch (MamaException e) {
    System.out.println (e);
}

/* Block until all objects have been destroyed. */ 
try {
    queue.destroyTimedWait (6000);
} catch (MamaException e) {
    Console.WriteLine ("Can't destroy queue: {0}\n", e.Message );
}

Additionally, the canDestroy function indicates if all objects using the queue have been destroyed:

/* Check if queue can be destroyed (no C++ API) */ 
mama_status status == mamaQueue_canDestroy (queue); 
if(status == MAMA_STATUS_QUEUE_OPEN_OBJECTS) 
{ 
    printf ("Queue cannot be destroyed as it is still in use.\n"); 
}
else if(status == MAMA_STATUS_OK) 
{
    printf (“Queue can be safely destroyed.\n”); 
}

/* Check if queue can be destroyed. */ 
mama_status status == mamaQueue_canDestroy (queue); 
if(status == MAMA_STATUS_QUEUE_OPEN_OBJECTS) 
{ 
    printf ("Queue cannot be destroyed as it is still in use.\n"); 
}
else if(status == MAMA_STATUS_OK) 
{
    printf (“Queue can be safely destroyed.\n”); 
}

// No Java API for this

/* Check if queue can be destroyed */ 
if (!queue.canDestroy())
    Console.WriteLine ("Can't destroy queue\n");
}

Destroying the Default Queue

Objects can be created on the default queue that are only destroyed when MAMA’s close function is called. The queue’s destroyTimedWait function is called internally by OpenMAMA and blocks for two seconds, after which a timeout error may be returned.

The two second time period can be increased by use of the following entry in the mama.properties file, expressed in milliseconds:

# Increase the native queue wait to 5 seconds 
mama.defaultqueue.timeout = 5000

Object Destroy Notifications

Ideally, applications will not attempt to destroy the queue until all the objects using it have been destroyed first. OpenMAMA can provide notification that an object has been fully destroyed by invoking a call-back function. The method of registration depends on the language being used.

Debugging Queue Destroy

If a programmer creates an object but forgets to destroy it, then the destroyWait function will block forever. Tracking down the offending object can be difficult for a large application and so OpenMAMA provides assistance via the object lock tracking property, which can be enabled in the mama.properties file.

    # Turn on queue tracking
    mama.queue.object_lock_tracking = 1

With this property turned on, OpenMAMA writes a log message every time the queue lock count is increased or decreased. The following output is an example of this message written at ‘NORMAL’ level.

2011-04-26 13:04:15:296: mamaQueue_incrementObjectCount(): queue 0x0092E600, owner 0x0092D480, new count 1. 
2011-04-26 13:04:15:296: mamaQueue_incrementObjectCount(): queue 0x0092E600, owner 0x00E65A20, new count 2. 
2011-04-26 13:04:15:296: mamaQueue_incrementObjectCount(): queue 0x0092E600, owner 0x00E65D30, new count 3. 
2011-04-26 13:04:15:296: mamaQueue_incrementObjectCount(): queue 0x0092E600, owner 0x00E6F488, new count 4. 
2011-04-26 13:04:15:296: mamaQueue_incrementObjectCount(): queue 0x0092E600, owner 0x00E65F00, new count 5.

Additionally, a block of memory is allocated at this point, which is freed whenever the lock count is decremented. This allows a memory leak detection tool to display the stack trace and pinpoint the location where the object was created. The following output is from valgrind running on linux and shows that a subscription was created in the subscribeToSymbols function that was not destroyed.

==15766== 8 bytes in 1 blocks are possibly lost in loss record 26 of 245 
==15766== at 0x4906795: calloc (vg_replace_malloc.c:418) 
==15766== by 0x4A72BED: mamaQueue_incrementObjectCount (queue.c:551) 
==15766== by 0x4A7A99A: mamaSubscription_setupBasic (subscription.c:649) 
==15766== by 0x4A7E562: mamaSubscription_setup (subscription.c:3238) 
==15766== by 0x4A7A119: mamaSubscription_create_ (subscription.c:349) 
==15766== by 0x4A7AC09: mamaSubscription_create (subscription.c:732) 
==15766== by 0x403BB8: subscribeToSymbols (in /var/userspace/gclarke/work/mama-50dev/
install/examples/mama/mamalistenc)
==15766== by 0x40379A: main (in /var/userspace/gclarke/work/mama-50-dev/install/ 
examples/mama/mamalistenc) 

Note The object lock tracking is turned off by default as it may cause degradation in application performance.

Dispatching

OpenMAMA provides the following mechanisms by which events can be dispatched from a queue:

1. Blocking Dispatching

   In this case a call to dispatch simply blocks, constantly dispatching events from the queue until stopDispatch is called. stopDispatch can be called from within a callback function/method from that queue or from another thread (which may be dispatching events from a different queue).

2. Timed Dispatching

   OpenMAMA provides a timedDispatch function, the behaviour of which is dependent on the underlying middleware being used. When using TIBCO Rendezvous the function behaves as tibrvQueue_timedDispatch whereby the function blocks until an event has been dispatched or the specified timeout interval has elapsed. When using 29West LBM or Data Fabric, the function only unblocks after the interval has elapsed regardless of events being dispatched or not.

3. Dispatch a single event

   The dispatchEvent function dispatches a single event from the queue and returns. If there are no events on the queue the function returns immediately.

Note Timed dispatching or single event dispatching are only available on queues created externally to the API. They are not available on the internal default event queue.

For full control over dispatching and threading in an application, you can create your own threads from which to dispatch on each queue created. Alternatively a Dispatcher can be used, which simply creates a new thread and starts dispatching (using the MAMA queue’s dispatch function) on the specified queue.

Queue Monitoring

The API provides the ability to receive notifications when certain conditions on an event queue occur. For instance, many application developers find it useful to know if the event queue is backing up, an early indication of a slow consuming client application.

The API currently supports registering for callbacks to be invoked when a high watermark for the queue is reached, that is, the number of outstanding events on the queue reaches a specified threshold, and for when the number of events returns to a low watermark.

Callbacks can be registered with a MamaEventQueue. These callbacks are invoked when certain conditions on the event queue are met. Which callbacks can be called and under what conditions they are called is middleware-dependent.

In all object orientated languages the callbacks are defined as concrete implementations of the MamaQueueMonitorCallback.

The high and low watermarks for an event queue are set via calls to setHighWatermark and setLowWatermark respectively. Callbacks for monitoring will not be called unless the watermark values have been set for a queue.

It is recommended that users specify a name for each queue being monitored. This will aid in logging and debugging the application.

If a monitoring feature is not available on a particular middleware an error will be returned, or an exception thrown.

//Callback class for receiving queue monitor events 
class QueueMonitorCallback : public MamaQueueMonitorCallback 
{
public:
    QueueMonitorCallback () {} 
    virtual ~QueueMonitorCallback () {} 
    virtual void onHighWatermarkExceeded (MamaQueue* queue, size_t size, void* closure)
    { 
        // Process the high watermark exceeded event
    }
    virtual void onLowWatermark (MamaQueue* queue, size_t size, void *closure) 
    { 
        // Process the low watermark event 
    }
} 
... 

queue->setQueueMonitorCallback (new QueueMonitorCallback (), "Closure"); 
queue->setHighWatermark (10000); 
queue->setLowWatermark (900);

static void MAMACALLTYPE
highWaterMarkCallback (mamaQueue     queue,
                       size_t        size,
                       void*         closure)
{
    // Process the high watermark exceeded event 
}

static void MAMACALLTYPE
lowWaterMarkCallback  (mamaQueue     queue,
                       size_t        size,
                       void*         closure)
{
    // Process the low watermark event 
}

mamaQueueMonitorCallbacks queueCallbacks;
queueCallbacks.onQueueHighWatermarkExceeded = highWaterMarkCallback;
queueCallbacks.onQueueLowWatermark  = lowWaterMarkCallback;
mamaQueue_setQueueMonitorCallbacks (queue, &queueCallbacks, "Closure");

queue.setQueueMonitorCallback (new MamaQueueMonitorCallback ()
{
    public void onHighWatermarkExceeded (MamaQueue queue,
                                         long      size)
    {
        // Process the high watermark exceeded event
    }

    public void onLowWatermark (MamaQueue queue,
                                long      size)
    {
        // Process the low watermark event
    }
});

//Callback class for receiving queue monitor events 
internal class ListenerQueueCallback : MamaQueueMonitorCallback
{
    public void HighWatermarkExceeded(MamaQueue mamaQueue, int size, object closure)
    {
        // Process the high watermark exceeded event
    }

    public void LowWatermark(MamaQueue mamaQueue, int size, object closure)
    {
        // Process the low watermark event
    }
}
... 

queue.setQueueMonitorCallback (new ListenerQueueCallback (), "Closure"); 
queue.setHighWatermark (10000); 
queue.setLowWatermark (900);

Event queue size

An application can find out the number of events currently on the event queue at any point, independent from the queue monitoring callbacks. This is available as getEventCount.

Queue groups

There is a further level of abstraction in the form of the MamaQueueGroup. The MamaQueueGroup, upon construction, creates the specified number of MamaQueue and MamaDispatcher objects and starts dispatching on the queues. The threading and dispatching is hidden from the application developers. The getNextQueue function returns the next available queue using round-robin, ensuring an even distribution of event sources across queues. For more control over dispatching of events within an application we recommend that you manipulate and manage queues directly by the application.

Developer Tips

1. Integrate an external event loop.

   The combined use of dispatchEvent and setEnqueueCallback can be used to integrate an OpenMAMA event loop into the event loop of another application when multi threading is being avoided. The setEnqueueCallback allows an application to be notified of enqueue events on a specified queue without having those events being dispatched. In the mamaQueueEnqueueCB callback function the application can post an event to its own event queue. When this event is dispatched the application then calls dispatchEvent. This strategy is sometimes employed in GUIs.

2. One to one relationship between queues and threads.

   For use of the API within the MAMA Advanced Publisher you have to maintain a one to one relationship between queues and dispatching threads. This is due to the internal use of message level sequence numbers used for data quality purposes. If dispatching from a single event queue coming from across multiple threads, the API can no longer enforce data quality correctly as messages cannot be guaranteed to be dispatched in the order they were received.

3. Minimize event queue growth.

   It is recommended that as little time as possible is spent executing code in the callback functions/ methods. Failure to do so can result in the event queue growing, ultimately running out of memory and/or message loss.

4. Using multiple queues/threads.

   There are several reasons why you may choose to use multiple threads/queues within an application built using the OpenMAMA API:

   • The application already uses an event loop to control processing. In this case the application cannot use the main thread to call MAMA’s start function. The alternative here is to run MAMA’s startBackground function which spawns a background thread on which to start dispatching on the default event queue or to integrate the two event queues (see point 1 above).
   • The application requires fine grained control over how events are dispatched. When using the default event queue developers have no control over how dispatching occurs. In order to leverage the further control provided by the queue object, developers must create their own event queues and dispatch from these on a separate thread. When this level of control is required it is recommended that the MamaQueueGroup is not utilized.

   • The nature of the application dictates that separate threads are used for processing data events.

     Consider the following scenario:

     One event source provides message updates which require a short processing time per message. A second event source provides message updates which require considerably more CPU cycles to process each message. To avoid this second source negatively effecting the processing of the first source of messages it may be beneficial to distribute the processing across two threads; one for each message source. Note: On a single CPU machine this approach is unlikely to provide much additional benefit.

When using multiple queues/dispatchers or the MamaQueueGroup utility class, developers need to be aware that the callback functions/methods can be called from multiple threads.

Subscriptions

Subscriptions in OpenMAMA provide the ability to register interest in a source of data for a specific symbol. The subscription interface hides the underlying middleware specific subscription concept (e.g. Listener when using the TIBCO Rendezvous middleware and Receiver when using the 29West LBM middleware). The OpenMAMA API is used to subscribe to market data from a market data source, e.g. the MAMA Advanced Publisher. The API also supports basic subscriptions, which are used to subscribe to data published using theOpenMAMA publishing functionality (see Section 13: Publishing for details). All sections of this chapter other than those dealing specifically with Basic Subscriptions, refer to market data subscriptions.

Note Feed handlers are used throughout this section to illustrate subscribing to market data. A MAMA Advanced Publisher can also be used.

Life Cycle of the MAMA Subscription

The OpenMAMA subscription moves through a number of different states during its lifetime, as defined in the following table.

The transitions between the various states are shown in the following state machine diagrams. The transitions also show the function calls that are required to change state.

State transitions for a Market Data Subscription

State transitions for a Basic Subscription

State transitions for activating and deactivating a Market Data Subscription

The complete state machine for a Market Data Subscription, including the main flow and the activation states

This lifecycle is reflected in the MamaSubscriptionState enumeration. The current state of the subscription can be obtained using the subscription’s getState function. This replaces the legacy isActive and isValid function calls which are now deprecated. These functions can still be used and will return ‘True’ if the subscription is in the Activated state.

Additional Notes:

1. If a function such as activate is called and the subscription is not able to make an appropriate transition, (as described by the state machine), then an error code of MAMA_STATUS_SUBSCRIPTION_INVALID_STATE will be returned.
2. Calling deactivate on a subscription in the allocated, setup, deactivated or destroyed state will have no effect and the function will simply return MAMA_STATUS_OK.
3. The unknown state indicates an error condition and should not normally occur.

Common Regular Subscription Behaviour

OpenMAMA provides a number of subscription types. All subscriptions share the following common concepts.

Symbol Namespace & Symbol

When a subscription is created, the user specifies a mandatory symbol namespace (e.g. “NASDAQ”) and a mandatory symbol (e.g. “MSFT”).

The symbol is the instrument identifier and should be unique when combined with the source. The NYSE TechnologiesMarket Data Platform Feed Handler Suite refers to this as the issue symbol.

The symbol namespace is a logical feed handler group identifier as configured by the feed handler administrators. It can be thought of as a namespace qualifier for market data and is useful in separating two publishers of data with the same symbology on the same transport settings.

For example, with a NASDAQ UTP feed handler configured with a namespace of “UTP” and the symbol “MSFT”, using the default UTP symbology on the feed handler, represents the identifier for the NBBO (National Best Bid and Offer) for Microsoft equities from NASDAQ.

Callbacks

Callbacks are registered with subscriptions when they are created. The callbacks are described the following table.

Regular Subscriptions

The sending of subscription requests to the feed handler, and the registering of interests with the underlying middleware are throttled when a regular subscription is created. This throttle is always controlled by the internal default queue in OpenMAMA, therefore, all subscription requests are sent by the thread calling MAMA’s start function.

When subscriptions are created using the default queue, onCreate() is guaranteed to be called before onMsg() and always from the thread calling start. However, if subscriptions are created on a user-created mamaQueue there is a possibility that onMsg() could get invoked before onCreate(). This is because onCreate() is called on the thread invoking MAMA’s start function whereas the onMsg() is called on the thread dispatching from the queue that was associated with the subscription upon creation.

Essentially, onCreate() is called as soon as the subscription request has been issued and the interest registered with the middleware. It is at this stage that the subscription creation has been picked off the throttle. For a user-created queue, because the data for the subscription is being dispatched on a separate thread, the possibility exists for the data to return from the publisher and be dispatched before the onCreate() callback gets scheduled to execute by the operating system thread scheduler.

This race condition exists as it was decided to keep the data callbacks lock-free for performance reasons. Regardless of which order the onCreate() and onMsg() callbacks are invoked, at the time of invocation the subscription can be treated as if it has been fully created and is considered valid in both cases.

Basic Subscriptions

Basic subscriptions are created immediately rather than being throttled, therefore all callbacks are invoked on the thread dispatching from the queue associated with the subscription.

Initial Images

By default, MAMA Advanced Publishers do not send all data for a symbol with every update. They only send modified data (deltas) or modified data along with additional specific static data. This saves on processing in the construction of messages and bandwidth when sending data over the network. Either option is more efficient than sending all data with every update.

While this is a very efficient mechanism, subscribing applications may have to wait an arbitrary length of time to obtain the latest value of all available published fields. To address this problem the OpenMAMA API provides the concept of initial images.

An initial image is a special message type that contains all the available fields in the market data publishers cache that are configured to be published. It is effectively a snapshot in time of the net effect of all updates on the instrument up to the point of subscription. As initial images contain all the fields in the publisher cache, the message size tends to be significantly larger than subsequent updates.

Initial images are identified as such through use of MamaMsgType’s typeForMsg() function. The MdMsgType reserved field stores this information within a message.

If this initial snapshot of data is not required, for example, for tick capture systems where only individual updates are of interest, the receipt of initial images on subscription can be disabled via a call to the subscription object’s setRequiresInitial(false) function.

In this case, the creation of a subscription simply informs the MAMA Advanced Publisher of a new subscriber and to start publishing for that instrument if not already doing so.

The rate at which initial images are sent can be controlled at the transport level.

Recaps

A recap is an initial image that is sent to the API to provide the client with the latest snapshot for a symbol in response to a data quality event on the infrastructure.

Recaps can be solicited, in that they are requested from the API in response to a sequence number gap being detected in the inbound messages for a subscription.

Conversely, recaps can be unsolicited, in that they are sent by the feed handlers under certain circumstances, for example, in the event of a cancel or correction being received by the MAMA Advanced Publisher.

The rate at which recaps are sent from the API can be controlled at the transport level.

Timeout/Retries

A MAMA Advanced Publisher may not respond to a subscription or initial image request in a timely fashion for a variety of reasons, such as network problems or being overloaded with requests. To further ensure receipt of an initial image within the API a subscription supports the concept of timeout intervals and a number of subscription retries before giving up and reporting a timeout error via the registered callbacks.

The number of retries, set via a call to MamaSubscription.setRetries(), specifies the number of attempts made to obtain an initial image for a subscription, with an interval defined by the timeout as described above.

The defaults here are three retries with a 10 second timeout between each retry.

Refreshes

A MAMA Advanced Publisher has no mechanism by which it can detect a subscribing client shutting down (i.e. In the case of an uncontrolled shutdown). Instead, there is the concept of refresh messages sent from the client to the MAMA Advanced Publisher to indicate that there is still some interest in the data being published. Refresh messages are sent once an hour, distributed over the hour, for each symbol a client has subscribed to. The MAMA Advanced Publisher stops publishing data for a symbol if it has not received a refresh message for that symbol during a defined, configurable, period of time (default 60 minutes).

To reduce the possibility of all clients sending refresh messages for the same symbols the MAMA Advanced Publisher sends a response to a refresh message to all clients. Upon receipt of this message each client puts the refreshed symbol on the end of its refresh list. Using this mechanism a MAMA Advanced Publisher will not be flooded with refresh messages when one is sufficient.

Throttling of Subscription Creation

By default, when OpenMAMA subscriptions are created, the subscription request is not sent immediately from the API. Instead, the request is placed on the default throttle queue to be sent at a later stage. Sending of subscription requests does not start until dispatching on the default event queue has commenced, i.e. by calling start(). This behaviour protects the messaging backbone from a storm of subscription requests and the MAMA Advanced Publisher from becoming overwhelmed with such requests.

The default throttling rate within the API, if none is specified, is 500 msg/sec. It is recommended that this value is significantly lowered in the following cases

• When using subscription types that result in larger initial values, e.g. Group and order book subscriptions.
• If there are a large number of subscribing applications that start up at the same time.
• The throttling is controlled at the transport level and applies to all subscriptions created on that transport.

Caching of Updates Prior to Initial

It is possible, on subscription creation, due to updates and initial values arriving along different communication paths, that the updates and initial values arrive out of sync. Consider the following scenario:

1. The MAMA Advanced Publisher is already publishing updates for a particular symbol (the client has subscribed after market open).
2. The client creates a subscription request.
3. The MAMA Advanced Publisher creates an initial image message from the current state of the cache for the subscribed symbol.
4. An update arrives to the MAMA Advanced Publisher for that symbol immediately after the initial value message has been created. The effects of this update on the MAMA Advanced Publisher symbol cache are published as normal.
5. The update from the previous step arrives at the client before the previously created initial value message. This could happen as the initial may be being published on TCP and the update on multicast.
6. The next update received at the client is not the next expected sequence number for that symbol. The client detects a gap and issues a recap request.

This scenario can lead to a number of recaps being send from a client resulting in an arbitrary amount of time before the client recovers. For instance if the scenario outlined occurred for an instrument that was not very liquid it may be some time before the feed handler receives and sends another update to the client.

To reduce the chance of this occurring the API can be configured to cache updates that arrive prior to receiving an initial image for a subscription. If the update subsequent to the initial value results in a gap being detected the API checks its message cache for the missing update before resorting to issuing a recap request.

The size of the cache used can be controlled on a per subscription basis using the subscription object’s setPreInitialCacheSize() function. The default is a cache of 10 messages to be stored prior to receiving the initial image.

Creating and Destroying Subscriptions

Subscriptions require that a transport, source and symbol are specified upon creation.

Subscriptions must be destroyed from within the subscriptions’ own callbacks or from other event callbacks on the same queue as the subscription callbacks are being dispatched.

class DisplayCallback : public MamaSubscriptionCallback 
{
public:
    DisplayCallback () {} 
    virtual ~DisplayCallback (void) {} 

    virtual void onCreate (MamaSubscription* subscription) 
    { 
        // The subscription has been created. 
    } 
    virtual void onError (MamaSubscription * subscription, 
                          const MamaStatus&  status, 
                          const char*        symbol) 
    { 
        // An error was encountered during the processing of the subscription 
    }
    virtual void onQuality (MamaSubscription* subscription, 
                            mamaQuality       quality, 
                            const char*       symbol) 
    { 
        // A data quality event has occurred. Likely to be a stale message 
        // due to a sequence number gap. 
    } 
    virtual void onMsg (MamaSubscription* subscription, 
                        const MamaMsg&    msg) 
    { 
        // Process the message 
    } 
    virtual void onGap (MamaSubscription* subscription) 
    { 
        // A gap has been detected for the subscription 
    } 
    virtual void onRecapRequest (MamaSubscription* subscription) 
    { 
        // A recap request has been sent for this subscription 
    }
}

.... 

DisplayCallback* callback = new DisplayCallback; 
MamaSubscription* subscription = new MamaSubscription (); 

// ... Set MamaSubscription properties if required 

// Create the source for the data -contains symbolnamespace and trasport
MamaSource* source = new MamaSource ("SourceId", transport, "NASDAQ");

subscription->create (queue,
                      callback, 
                      source, 
                      "MSFT", 
                      "My Closure");

....
// Destroy the subscriptionsubscription->destroy ();
delete subscription; 

static void subscriptonOnMsg (mamaSubscription subscription,
                              mamaMsg msg,
                              void* closure,
                              void* itemClosure)
{
    // Process the message
}

static void subscriptionOnError (mamaSubscription subscription,
                                 mamaStatus status,
                                 void* platformError,
                                 const char* subject,
                                 void* closure)
{
    // Handle error condition
}

static void subscriptionOnCreate (mamaSubscription subscription,
                                  void* closure)
{
    // The subscription has been created
}

static void subscriptionOnQuality (mamaSubscription subsc,
                                   mamaQuality quality,
                                   const char* symbol,
                                   void* closure)
{
    // Handle change in data quality update
}

static void subscriptionOnRecapRequest (mamaSubscription subsc,
                                        void* closure)
{
    // A recap has been sent for the subscription
}

static void subscriptionOnGap (mamaSubscription subsc,
                               void* closure)
{
    // A gap has been detected for this subscription
}

...

mamaMsgCallbacks callbacks;
callbacks.onCreate = subscriptionOnCreate;
callbacks.onError = subscriptionOnError;
callbacks.onQuality = subscriptionOnQuality;
callbacks.onMsg = subscriptonOnMsg;
callbacks.onGap = subscriptonOnGap;
callbacks.onRecapRequest = subscriptonOnRecapRequest;
...
/*Create the source for the data - contains symbolnamespace and trasport*/
mamaSource subscriptionSource = NULL;

mamaSource_create (&subscriptionSource)
mamaSource_setId (subscriptionSource, "SourceForSubscription");
mamaSource_setTransport (subscriptionSource, transport);
mamaSource_setSymbolNamespace (subscriptionSource, "NASDAQ");

...

mamaSubscription subscription = NULL;
mamaSubscription_allocate (&subscription);

// ... Set mamaSubscription properties

mamaSubscription_create (subscription,
                         queue,
                         &callbacks,
                         subscriptionSource,
                         "symbol",
                         "My Closure");
....
//Destroy the subscription
mamaSubscription_destroy (subscription);
mamaSubscription_deallocate (subscription);

class DisplayCallback implements MamaSubscriptionCallback
{
    public DisplayCallback () {}
    public void onCreate (MamaSubscription subscription)
    {
        // The subscription has been created
    }
    public void onError (MamaSubscription subscription,
                         short wombatStatus,
                         int platformError,
                         String subject,
                         Exception e)
    {
        // An error has occurred during subscription processing
    }
    public void onQuality (MamaSubscription subscription,
                           short quality)
    {
        // A data quality event has occurred.
    }
    public void onMsg (MamaSubscription subscription,
                       MamaMsg msg)
    {
        // Process the message
    }
    public void onRecapRequest (MamaSubscription subscription)
    {
        // A recap request has been sent for this subscription
    }
    public void onGap (MamaSubscription subscription)
    {
        // A gap has been detected for this subscription
    }
    public void onDestroy (MamaSubscription subscription)
    {
        // The dubscription has been destroyed
    }
}

DisplayCallback callback = new DisplayCallback ();
MamaSubscription subscription = new MamaSubscription ();

//set MamaSubscription properties

MamaSource source = new MamaSource ("SourceId", transport, "NASDAQ");

subscription.create (queue,
                     callback,
                     source,
                     "MSFT",
                     "My Closure");
....

//Destroy the subscription
subscription.destroy ();

class DisplayCallback : MamaSubscriptionCallback
{
    public DisplayCallback () {}
    public void onCreate (MamaSubscription subscription)
    {
        // The subscription has been created
    }
    public void onError (MamaSubscription subscription,
                         MamaStatus.mamaStatus status,
                         String subject)
    {
        // An error has occurred during subscription processing
    }
    public void onQuality (MamaSubscription subscription,
                           mamaQuality quality)
    {
        // A data quality event has occurred.
    }
    public void onMsg (MamaSubscription subscription,
                       MamaMsg msg)
    {
        // Process the message
    }
    public void onRecapRequest (MamaSubscription subscription)
    {
        // A recap request has been sent for this subscription
    }
    public void onGap (MamaSubscription subscription)
    {
        // A gap has been detected for this subscription
    }
    public void onDestroy (MamaSubscription subscription)
    {
        // The dubscription has been destroyed
    }
}

DisplayCallback callback = new DisplayCallback ();
MamaSubscription subscription = new MamaSubscription ();

//... Set MamaSubscription properties

MamaSource source = new MamaSource ("SourceId", transport, "NASDAQ");
subscription.create (queue,
                     callback,
                     source,
                     "MSFT",
                     "My Closure");
....

//Destroy the subscription
subscription.destroy ();

Note In the above example the call to create() is the equivalent of calling setup() followed by activate() on a subscription.

Reusing Subscriptions

It is possible to reuse already created subscriptions within the C++ version of the API. A subscription can be destroyed and subsequently recreated, or it can be deactivated and subsequently reactivated.

When a subscription is deactivated, all state objects created in the call to setup(), and associated with the subscribed symbol, are retained. The act of deactivating simply deregisters interest in the symbol with the underlying middleware. No more updates for the subscription are received after deactivate is called. Reactivating a subscription results in an initial value request being sent, if the subscription is using initial values, and interest in the symbol being registered with the underlying middleware. In this case the same symbol, symbolnamespace, transport and queue are used when the subscription is activated.

If reuse of the subscription for a different symbol is required the subscription can be destroyed. This tears down all state objects within the subscription that associate it with a particular symbol. The subscription can be recreated, using setup() or activate(), with completely new attributes enabling it to subscribe to a separate symbol.

Subscription Types

There are a number of different types of subscriptions that can be created using OpenMAMA, depending on the nature of the data being subscribed to. The type of subscription being created and its behaviour are controlled via two properties of the subscription that can be specified at creation time. These are:

• SubscriptionType:
  • MAMA_SUBSC_TYPE_NORMAL: Regular market data subscription
  • MAMA_SUBSC_TYPE_BOOK: Order Book Subscription
  • MAMA_SUBSC_TYPE_GROUP: Group subscription
  • MAMA_SUBSC_TYPE_BASIC: Basic Subscription
  • MAMA_SUBSC_TYPE_DICTIONARY: Dictionary Subscription
  • MAMA_SUBSC_TYPE_SYMBOL_LIST: Symbol List Subscription
  • MAMA_SUBSC_TYPE_SYMBOL_LIST_BOOK: Book Symbol List Subscription
• ServiceLevel: The additional behavior for the specific type of subscription. Currently only two of the values are supported:
  • MAMA_SERVICE_LEVEL_REAL_TIME: A real time subscription receives updates when they occur.
  • MAMA_SERVICE_LEVEL_SNAPSHOT: A snapshot subscription receives only an initial value and no subsequent updates.

The default values for these, if not explicitly specified upon subscription creation, are MAMA_SUBSC_TYPE_NORMAL and MAMA_SERVICE_LEVEL_REAL_TIME resulting in a real time market data subscription.

The behaviour of each of the subscription types below can be further qualified through the use of the service level.

Basic Subscriptions

A basic subscription allows users of the OpenMAMA API to subscribe to non-market data that is being published via an OpenMAMA based publisher. A basic subscription is created by specifying a subscription type of MAMA_SUBSC_TYPE_BASIC. As initial values are a concept specific to the MAMA Advanced Publisher, the only value of service level which is supported for basic subscriptions is MAMA_SERVICE_LEVEL_REAL_TIME (the default value when creating a subscription). A basic subscription does not provide initial values, recaps, data quality or refreshes. Throttling of subscription requests, however, is supported.

Entitlements

When subscribing to market data, and when that data is configured to include entitle codes, the API enforces entitlements as defined in the entitlements server. Entitlements are not enforced for ‘basic’ subscriptions.

The OpenMAMA API obtains its user based entitlements on application startup. The entitlements for the logged on application user are obtained from a HTTP server. The API obtains its list of available entitlements servers from the mama.properties file. This occurs on the initial call to open. The following example shows how to specify the location of entitlement servers using a comma separated list.

entitlement.servers=server1:9090,server2:8080 

The API checks each server, using round-robin, to locate the entitlements for the current user. If no entitlements servers are present, or no entitlements exist for the current user, the call to open() fails, with a status of ENTITLE_NO_SERVERS_SPECIFIED or ENTITLE_NO_USER, respectively.

Entitlements are enforced at two points during the subscription process:

1. At the point of subscription creation.

   When a subscription is being created the API checks the entitlement rules for the user to determine whether there are sufficient privileges to allow the user to create the subscription. This check is based on the symbol string being subscribed to. If the user is not entitled to subscribe to the symbol in question, the call to create() returns a value of MAMA_STATUS_NOT_ENTITLED.

2. On receipt of the first update to the client.

   The user must also be entitled to view data with a specific injected entitle code. When the API gets the first message for a subscription it checks if the user is entitled to view data with the specific entitle code. If the user is not entitled to this code the onError() subscription callback is invoked with an error code of MAMA_STATUS_NOT_ENTITLED.

Note If a message with an injected entitle code is received by a basic subscription it will be rejected, the assumption being that this is a market data message and should not be available to basic data subscribers.

Threading

This section identifies the threads on which the various callbacks within the API are invoked. In the majority of cases the threading model used is the same across language variations and middlewares. Where the API deviates from this, details, and an explanation, are provided.

Note The “Default Queue” is the thread that invokes start(), and therefore the thread that is dispatching on the internal default event queue.

The “Dispatch Queue” is the thread currently dispatching on the associated event queue. This can be the default queue or a user created event queue (MamaEventQueue).

The following callbacks are those that are associated with a MamaSubscription.

MamaSubscription Callbacks

Notes:

• The onCreate() callback is always called from the thread dispatching on the default queue. This is because subscriptions are centrally throttled on a per transport basis. The effect is when creating subscriptions using an OpenMAMA queue other than the default queue, it is possible that the onMsg() callback may get invoked prior to the invocation of the onCreate() callback. This behavior is being maintained as it is preferable to the overhead and latency that would be incurred in trying to synchronize the two.
• The majority of invocations of onQuality() will be called from the thread dispatching on the subscriptions associated event queue. When a subscription state is set to stale as a result of a detected gap, the onQuality() callback is always invoked from the dispatch queue.

MamaTimer Callbacks

MamaIo Callbacks

MamaTransport Callbacks

Event Queue Callbacks

MamaDqPublisher Callbacks

OpenMAMA Dictionary

The Data Dictionary is a data structure, obtained from the advanced publisher, which provides a mapping between field identifiers (FID’s) and field names for a superset of all fields that can be sent on the platform. It also provides data type information for each of the fields.

Creating a data dictionary in OpenMAMA is similar to creating a subscription, as the data dictionary request is a specialized form of subscription.

Creating the Data Dictionary (from platform)

class DictionaryCallback : public MamaDictionaryCallback 
{
public:
    DictionaryCallback () {} 
    virtual ~DictionaryCallback () {} 
    virtual void onTimeout (void) 
    { 
        // Could not get dictionary. The timeout interval of 60 seconds has elapsed. 
    } 
    virtual void onError (const char* errMsg) 
    { 
        // An error was encountered fetching the dictionary. 
    } 
    virtual void onComplete (void) 
    { 
        // The dictionary has been successfully fetched. 
    }
}

....

DictionaryCallback callback; 
MamaDictionary*    dict = new MamaDictionary;

dict->create (transport, 
              queue,
              &callback, 
              "WOMBAT"); 

static void dictOnTimeout (mamaDictionary dictionary,
                           void* closure)
{
    // Could not get dictionary. The timeout interval of 60 seconds has elapsed.
}

static void dictOnError (mamaDictionary dictionary,
                         const char* errMsg,
                         void* closure)
{
    // An error was encountered fetching the dictionary.
}

static void dictOnComplete (mamaDictionary dictionary,
                            void* closure)
{
    // The dictionary has been successfully fetched.
}

....

mamaDictionary dict = NULL;

mamaDictionaryCallbackSet callbacks;
callbacks.onComplete = dictOnComplete;
callbacks.onTimeout = dictOnTimeout;
callbacks.onError = dictOnError;

mama_createDictionary (&dict,
                       callbacks,
                       transport,
                       queue,
                       "WOMBAT",
                       closure);

boolean dictionaryComplete = false;
...
class DictionaryCallback implements MamaDictionaryCallback
{
    public synchronized void onTimeout ()
    {
        // Timedout fetching dictionary
        notifyAll ();
    }

    public synchronized void onError (final String errMsg)
    {
        // An error was encountered fetching the dictionary.
        notifyAll ();
    }

    public synchronized void onComplete ()
    {
        dictionaryComplete = true;
        notifyAll ();
        // The dictionary has been successfully fetched.
    }
}

...

DictionaryCallback callback = new MamaDictionaryCallback ();
MamaDictionary dictionary = null;

synchronized (callback)
{
    MamaSubscription subscription = new MamaSubscription ();

    dictionary = subscription.createDictionarySubscription (
                        callback,
                        transport,
                        "WOMBAT");
    callback.wait (30000);

    if (!dictionaryComplete)
    {
        System.err.println ("Could not retrieve dictionary");
        System.exit (1);
    }
}

....

// Dictionary retrieval has been successful. Process subscriptions etc.

class DictionaryCallback : public MamaDictionaryCallback
{
    virtual void onTimeout (void)
    {
        // Could not get dictionary. The timeout interval of 60 seconds has elapsed.
    }
    virtual void onError (const char* errMsg)
    {
        // An error was encountered fetching the dictionary.
    }
    virtual void onComplete (void)
    {
        // The dictionary has been successfully fetched.
    }
}

....

DictionaryCallback callback;
MamaDictionary* dict = new MamaDictionary ();

dict.create (queue,
             callback,
             "WOMBAT",
             3,
             10.0);

The dictionary request has successfully completed and the mamaDictionary is available for use once the onComplete() callback function/method has been invoked.

If there has been no response to the data dictionary request, the onTimeout() callback function/ method is invoked after 60 seconds has elapsed.

If any other error was encountered during the processing of the request, the onError() callback function/method is invoked passing back the appropriate error status.

The data dictionary can be obtained from a single MAMA Advanced Publisher instance. However, it is more typical that a dedicated dictionary serving process (e.g. mamadict) is running on the network that provides a superset of all the fields being used across all the market data publishers on the market data backbone.

The default source value for data dictionary retrieval is “WOMBAT”. This value is configurable for the dedicated dictionary publisher.

Using the Data Dictionary

The OpenMAMA representation of the Data Dictionary comprises a number of MamaFieldDescriptor objects, one for each field in the dictionary. As market data messages on the OpenMAMA platform generally only contain the FID, the dictionary can be interrogated at run time to find a full description of a message field. Not sending the field name is an optimization when sending messages, saving CPU processing time in the construction of the messages, and bandwidth by reducing the size of the messages.

The MamaFieldDescriptor can be obtained in three ways from the dictionary: by FID, by name, or through iteration across all field descriptors. For example, if the FID for a field is available the field descriptor for the field can be obtained using the dictionary’s getFieldByFid() function.

Once the MamaFieldDescriptor has been obtained, the field details can be accessed using the functions:

• getName()
• getFid()
• getType()
• getTypeName()

Developer Tips

The data dictionary can be serialized to and from a mamaMsg (the dictionary is received from the advanced publisher as a MamaMsg). The dictionary supports the following two operations to facilitate this:

• getDictionaryMessage()
• buildDictionaryFromMessage()

Note A new message is allocated for each invocation of getDictionaryMessage(). It is the responsibility of the caller to destroy the message(s) when no longer required in runtimes without garbage collection.

Once the underlying MAMA Message for a dictionary has been obtained, the message bytes can be written to file. The message can then be later reconstructed from the bytes in the file and from this message the dictionary can be recreated. Memory for the new dictionary is allocated through the dictionary’s create function.

Deserializing the data dictionary

MamaDictionary* dictionary = new MamaDictionary;
dictionary->buildDictionaryFromMessage (message); 

mamaDictionary dictionary = NULL;
mamaDictionary_create (&dictionary);
mamaDictionary_buildDictionaryFromMesage (dictionary, message);

MamaDictionary dictionary = new MamaDictionary();
dictionary.buildDictionaryFromMessage (message); 

MamaDictionary dictionary = new MamaDictionary();
dictionary.buildDictionaryFromMessage (message); 

Messages

The MAMA Message abstracted interface provides a wrapper for the underlying wire message formats supported on a particular messaging middleware.

Accessing Data

The MAMA Message object supports direct field access through a suite of strongly typed accessor functions/ methods. For example, getI8() or getF64().

A scalar field can be obtained through an accessor for a type larger than the one being accessed, when the larger type can hold the smaller without loss of precision. For instance, getI32() can be used to get fields of type U16, I16, U8 etc.

In the object oriented APIs, each of the accessor methods are overloaded with a version that accepts a MamaFieldDescriptor instead of the name/FID combination. It is recommended that applications use this variation of the accessor when accessing field data. A representation of the field data as a string can also be obtained via getFieldAsString(). However, this is obviously less efficient than using the correct type accessor for the field.

When accessing string fields using the getString function, the result points to a reference of the string held internally in the MAMA Message object. This memory is owned by the object and does not need to be explicitly freed.

If one of the strongly typed accessors is called on a MAMA Message and the field is not found, the function returns a mamaStatus value of MAMA_STATUS_NOT_FOUND, or an exception is thrown.

Message Creation

Applications need to create their own MAMA Messages when using the publishing capability of the API.

A MAMA Message can be created in a number of ways. The default create() creates a message with the default payload bridge createForPayload() creates a message using the specified payload. A message can be recreated from a byte buffer using createFromByteBuffer().

Note The OpenMAMA API has an internal list of reserved fields used for passing message header information and other data. It is strongly recommended that users of the API do not use FIDs of 100 or lower, or the field 496 if using entitled APIs, as these are used to describe the internal reserved OpenMAMA fields.

Note The FID uniquely identifies a field within a message, not the FID/name combination. This is an important distinction as the name is only used to search for fields when a field with the specified FID is not found.

Message creation

MamaMsg msg;
msg.create ();
msg.createForPayload (MAMA_PAYLOAD_QPID);

mamaMsg msg = NULL;
mamaMsg_create (&msg);
mamaMsg_createForPayload (&msg, MAMA_PAYLOAD_QPID);

MamaMsg msg = new MamaMsg (MamaPayloadType.MAMA_PAYLOAD_QPID);

MamaMsg msg = new MamaMsg (MamaPayloadType.MAMA_PAYLOAD_QPID);

Fields can be added to messages using the individually typed mutator functions available. When adding fields to messages both the field name and the FID can be specified.

Field Iteration

OpenMAMA enables iteration through all fields in a message. An application can pass a callback to iterateFields() which is invoked for each field in the message. The callback function/ method provides access to a MamaMsgField object. The field type can be obtained from the MamaMsgField object and the appropriate accessor can be invoked. All typed accessor functions or methods for the MamaMsg are also available for the MamaMsgField object. The field also supports obtaining the field data as a string using getAsString(). This approach is less efficient than strong typed access, in the same way as when obtaining stringified data directly from a message.

Field iteration

class MsgIteratorCallback : public MamaMsgFieldIterator 
{
public:
    MsgIteratorCallback () {} 
    virtual ~MsgIteratorCallback (void) {} 
    virtual void onField (const MamaMsg&      msg, 
                          const MamaMsgField& field, 
                          void*               closure) 
    { 
        //process the data in the MamaMsgField 
    }
}

.... 

MsgIteratorCallback callback; 
msg->iterateFields (callback,
                    dictionary,
                    "My Closure");

static void msgOnField (const mamaMsg msg,
                        const mamaMsgField field,
                        void* closure)
{
    // Process the data in the mamaMsgField.
}

....

mamaMsg_iterateFields (msg,
                       msgOnField,
                       dictionary,
                       "My Closure");

class MsgIteratorCallback extends MamaMsgFieldIterator
{
    public void onField (MamaMsg msg,
                         MamaMsgField field,
                         MamaDictionary dict,
                         Object closure)
    {
        // Process the data in the MamaMsgField
    }
}
....
msg.iterateFields (new MsgIteratorCallback (),
                   dictionary,
                   "My Closure");

class MsgIteratorCallback : MamaMsgFieldIterator
{
    public void onField (MamaMsg msg,
                         MamaMsgField field,
                         object closure)
    {
        // Process the data in the MamaMsgField
    }
}

....

msg.iterateFields (new MsgIteratorCallback (),
                   dictionary,
                   "My Closure");

There is also a separate iterator implementation as an alternative to the callback method. Using this the iterator points to a particular MamaMsgField object, and can be incremented to the next field by the user. A NULL is returned after the last field has been returned. It is possible to reset the iterator to the start of the message at any time.

Field iteration without callback

MamaMsgIterator iterator (myMamaListen->getMamaDictionary ());
msg.begin(iterator);
while (*iterator != NULL)
{
    // Process the data in the MamaMsgField
    ++iterator;
}

mamaMsgIterator iterator;
mamaMsgField field;

mamaMsgIterator_create (gDictionary, &iterator);
if (mamaMsgIterator_associate(msg, iterator) != MAMA_STATUS_OK)
{
    // Failed to associate iterator with message
}
else
{
    mamaMsgIterator_begin(iterator);
    while ((field = mamaMsgIterator_next(iterator)) != NULL)
    {
        // Process the data in the MamaMsgField
    }
}

for (Iterator iterator = msg.iterator(dictionary); iterator.hasNext();)
{
    MamaMsgField field = (MamaMsgField) iterator.next();
    // Process the data in the MamaMsgField
}

MamaMsgIterator iterator = new MamaMsgIterator (myMamaListen->getMamaDictionary ());
MamaMsgField field = null;

msg.begin (ref msgIterator);

while ((field = iterator.getField ()) != null)
{
    // Process the data in the MamaMsgField
    iterator++;
}

Each message can only have one iterator associated with it, though the same iterator can be used for more than one message.

Special Data Types

Some data types supported by a mamaMsg are specific to the OpenMAMA API.

MamaDateTime

A date/time representation with additional hints for precision, advanced output formatting and support for time zone conversion (using the MamaTimeZone type).

Hints include:

• Whether the time stamp contains a date part, a time part, or both.
• The level of accuracy (if known) of the time part e.g., minutes, seconds, milliseconds, etc.

The output format strings are similar to that available for the strftime() function, plus:

• %; adds an optional (non-zero) fractional second to the string
• %: adds fractional seconds based on the accuracy hint (including trailing zeros, if the accuracy hint indicates it should).

The following table provides examples of output.

MamaPrice

MamaPrice is a special data type for representing floating point numbers that often require special formatting for display purposes, such as prices. MamaPrice contains the 64-bit (double precision) floating point value and an optional display hint.

A MamaPrice may be marked as valid or invalid. A valid MamaPrice is one that contains a currently valid value. This can be used to differentiate between a zero (valid) value and an absence of value for example. A MamaPrice is set to valid by default when it is created with a value, or a value is explicitly set.

The set of display hints includes hints for:

• The number of decimal places
• The fractional denominator, to a power of two
• Special denominators used in the finance industry e.g., halves of 32nds

Developer Tips

When using the WombatMsg wire format and C++, it is more efficient to use field iteration rather than direct random field access. This results in significant performance improvements in the C++ OpenMAMDA API.

The underlying message can be written to a byte buffer, which can be serialized to file. A message can be created from an existing byte buffer. The mamaMsg determines the underlying message format from the provided byte buffer and constructs the message accordingly.

Note The buffer returned is not a copy and therefore should not be altered once obtained. To do so can corrupt the message.

Using a byte buffer

byte[] buffer = mSecondMessage.getByteBuffer();

const char* buffer = NULL; 
mama_size_t bufferSize = 0; 
MamaMsg newMessage = new MamaMsg; 
msg.getByteBuffer (&buffer,&bufferSize) 
//write the buffer to file... 
//read the buffer from file... 
msg.createFromBuffer (buffer, bufferSize);

const char* buffer = NULL;
mama_size_t bufferSize = 0;
mamaMsg newMessage = NULL;
mamaMsg_getByteBuffer (msg,(const void**)&buffer,&bufferSize);
//write the buffer to file...
//read the buffer from file...
mamaMsg_createFromByteBuffer (&newMessage,(const void*)buffer,bufferSize);

byte[] buffer = mSecondMessage.getByteBuffer();
// write the buffer to file...
// read the buffer from file...
mMessage.createFromByteBuffer (buffer);

byte[] buffer = null;
int size = 0;
msg.getByteBuffer (ref buffer, ref size);
// write the buffer to file...
// read the buffer from file...
msg.createFromBuffer (buffer, size);

Obtain your MamaFieldDescriptor based on the field name from the data dictionary on application startup. Cache the MamaFieldDescriptors and use them for field access later on, removing the need to know the actual FIDs for individual fields in messages.

Do not delete messages received in subscription callbacks. The API reuses message instances for performance reasons. If a message is required to live beyond the scope of a callback use copy() to create a deep copy of the message. Alternatively detach() transfers ownership of the message from the API to the caller of the function. In this case it is the responsibility of the calling application to destroy the message when it is no longer needed. Similarly, when extracting submessages/ arrays of sub-messages from a message, the memory for these is freed/deleted when the parent message goes out of scope in a callback or is deleted if owned by the application using the API.

MamaMsg Wire Format Conversion Matrix

This table shows the recommended possible data type conversions from MamaMsg wire format types when extracting data from a MamaMsg field that are supported across all middlewares and message types.

Note Other conversions may be possible depending on the middleware and message type being used.

Data Quality

The OpenMAMA API provides a number of features to ensure the integrity of the inbound data when subscribing to market data.

Note This does not apply to ‘basic’, non market data, subscriptions.

Messages are sent with an injected sequence number field (MdSeqNum, FID 10) which contains a sequence number for each symbol (each individual symbol within a group subscription has its own sequence number). When a gap is detected in this sequence number OpenMAMA marks the message as being stale (STATUS_STALE) before passing the message to the client application. The API requests a recap image for the symbol in question. Once the recap has been received the API resets the internal expected sequence number and marks all subsequent messages as being OK (STATUS_OK).

Recaps are generally solicited from a client when a sequence number gap is detected. However, after a fault tolerant takeover the new primary may send unsolicited recaps for all instruments that have changed during a configurable interval. In this case the receipt of the recap message is not preceded with a gap callback. Recaps are sent using broadcast from the MAMA Advanced Publisher. Therefore, a client may receive an unsolicited recap as a result of another client’s request. Again, this recap is not preceded with a gap callback notification. In each of these cases, all clients subscribed to the recapped instrument on that transport receive the recap.

The sending of recap requests is controlled in the same manner as initial value requests. A configurable number of recap requests can be sent across a specified interval. If a recap, after the specified number of attempts have been made, is not received, the onError() callback is invoked with a status of MAMA_STATUS_TIMEOUT. Under normal stable operating conditions a timeout for a recap should not occur. If encountered it is generally an indication of problems with the client and/or environment. The recommended action is to recreate the subscription (the abnormal condition possibly being transient) and alert administrator(s) to a possible application/environment issue. This course of action applies equally to recap timeouts for group subscriptions.

In the API, when a sequence number gap is detected, the subscription onQuality() callback is invoked with a value of MAMA_QUALITY_STALE. Once the condition has been resolved (a recap received), the onQuality() callback is invoked once again with a value of MAMA_QUALITY_OK. For specific languages, see the MamaQuality enumerated type.

Subscription data quality events can be captured by implementing the onGap() and onRecapRequest() subscription level callbacks.

The following describes the series of events that occur within the API when a gap is detected:

1. OpenMAMA detects a sequence number gap for an instrument.
2. The onGap() callback is invoked.
3. The subscription onRecapRequest() callback is invoked. A recap is requested if the subscription does not currently have a status of STALE and is not already waiting for the response to a recap request.
4. As this is a quality change from OK to STALE the onQuality() callback for the subscription is invoked with a status of STALE.
5. The message is passed up to the application with a status of STALE.
6. If any updates are received prior to the recap being received, the onGap() callback is invoked and the message is passed to the application with a status of STALE. This does not result in onQuality() being invoked, as this gap does not result in a quality state change, or a recap being requested, as the API has already sent a request and is waiting for a response.
7. The requested recap arrives.
8. The onQuality() callback is invoked with a status of OK.
9. The message is passed to the application.

Sequence number checking can be disabled in the API if required. Calling MamaSubscription’s setRecoverGaps() informs the API to no longer check the injected sequence number for gaps.

Sequence number gaps can be indicative of a number of possible problems:

• Data loss on the inbound OS socket buffers (client machine -multicast).

  This can result from the client application monopolizing the CPU for a prolonged period of time thereby being unable to consume data from the inbound socket buffer in a timely fashion.

  Increasing the OS inbound socket buffer size can help to address this problem. The increased buffer size helps deal with short lived events that consume all CPU cycles. However, if the client continues to consume all CPU cycles, increasing the socket buffer sizes only delays the inevitable.

  This problem typically occurs with multicast data:

  • For TIBCO Rendezvous based clients, it is the CPU of the Rendezvous Daemon (RVD) that is of concern.
  • For 29West LBM based clients, it is the CPU of the application using the 29West LBM libraries that is of concern.

• Data loss on the outbound socket buffers (sending machine -TCP).

  This can result from a client application being unable to consume inbound data as quickly as it is being sent. The likely cause of this is the client application using all CPU cycles for a period of time.

  Unlike multicast, which sends as fast as possible regardless of the ability of the client to keep up, TCP enforces flow control if a client cannot process messages in a timely fashion and buffers data until such time as the client can process again. If this condition persists the sender buffers will ultimately overflow and lose data.

  This occurs with 29west LBM TCP and communication between an RVD and the TIBCO Rendezvous API.

  The ability for a client application to consume data from an incoming socket buffer can be impacted by a number of sources of resource contention. For example:

  • Running out of physical memory resulting in significant paging.
  • High levels of context switching on an overloaded machine -e.g. very high load average on Linux.
  • Significant IO blocking.
  • Lack of available CPU cycles.

• Data loss on the network

  Various network conditions can result in lost data. Such as:

  • Unreliable cabling
  • Mis-configured switches
  • Receive buffer sizes too small
  • Mis-configured NIC’s
  • Router buffer overflow
  • Congested network

The OpenMAMA API also attempts to monitor the health of the underlying messaging infrastructure. Upon detecting a problem like a failed transport, an overflowing queue or a slow consumer, the API marks all inbound messages as ‘possibly stale’ (MAMA_QUALITY_MAYBE_STALE) while the condition persists. OpenMAMA continues to mark messages as STATUS_STALE while the condition persists and does not request a recap for the subscriptions until the problem subsides.

Note Messaging infrastructure level checking is only currently available on the TIBCO Rendezvous platform

Note The behavior of data quality in OpenMAMA does not vary with middleware.

Data Quality for Group Subscriptions

Items within a group subscription have their own individually tracked sequence number. As such, data quality for groups is tracked on a per item basis. On detecting a gap for an item within a group subscription a recap is requested for that item only.

Data Quality and Fault Tolerant Takeovers

The MAMA Advanced Publishers can be configured to run as primary/secondary pairs in hot/hot fault tolerant mode. Any any point in time only the primary feed is publishing data. When a fault tolerant event occurs, a secondary publisher assumes the role of primary and all instruments that have ticked during a configurable period longer than the fault tolerant takeover interval are recapped by the new primary. It is then assumed that data for all remaining instruments are in sync with the last message sent by the previous primary.

The OpenMAMA API tracks the senderId, a 64 bit field that uniquely identifies a publisher on the platform (MamaReservedFieldSenderId, FID 20), if present. A change in the sender Id is taken to represent a fault tolerant event at the publisher level. The current behaviour is to assume that an update with a different senderId is the next expected update, and to reset the internal data quality state with the senderId and sequence number of the message, marking the subscription quality as OK if required. Unsolicited recaps will have been sent from the new primary feed on any symbols for which all updates may not have been received.

Publishing

Although the OpenMAMA API is primarily used within subscribing applications, it can also be used for publishing. Data can be published on a particular symbol/topic, in either a request response paradigm or the normal publish-subscribe environment. Two types of publishing are available: basic and advanced.

Basic publishing is the counterpart to basic subscriptions. Basic subscriptions do not require initials, seqnums, refreshes and so on, and basic publishing does not cater for these either. It allows the sending of simple messages of any fields with or without field names or FIDs. Basic publishing is recommended in a generic messaging environment or for admin and control messages between market data applications.

Advanced publishing is a much stricter framework that is designed to provide the necessary tools for building a full market data publishing source for fully interacting with OpenMAMA clients. The advanced publishing classes build on the basic publisher but still use the same underlying concepts.

Basic Publishing

Data is sent from a MamaPublisher object. On creation of a publisher, a transport and an outbound topic are specified. The outbound topic (“MY_PUB_TOPIC”) is the symbol or topic for which a subscribing application must create a basic subscription for.

MamaPublisher

MamaPublisher publisher = new MamaPublisher; 
publisher->create (transport, "MY_PUB_TOPIC");

mamaPublisher publisher = null;
mamaPublisher_create (&publisher,
                      transport,
                      "MY_PUB_TOPIC",
                      NULL, /* Not needed for basic publishers */
                      NULL  /* Not needed for basic publishers */
);

MamaPublisher publisher = new MamaPublisher ();
publisher.create (transport, "MY_PUB_TOPIC");

MamaPublisher publisher = new MamaPublisher ();
publisher.create (transport, "MY_PUB_TOPIC");

Sending a message from a publisher is straightforward: use the send() method, passing in the MamaMsg object to be sent. A message published in this fashion is sent from OpenMAMA immediately. However, depending on the middleware configuration, there may be a short delay before its actually sent. The message is normally sent from the thread calling send(), but there are some exceptions to this depending on middleware configuration.

publisher->send (msg);

mamaPublisher_send (publisher, msg);

publisher.send (msg);

publisher.send (msg);

If there is a need to control the publish rate of messages, use sendWithThrottle(). This places the message or action on the internal throttle queue to be sent at some time in the future. The throttle rate of the transport, on which the publisher was created, is used. The default rate is 500 per second. When the message is sent, a complete callback will be called which is passed into the sendWithThrottle() method. It is the responsibility of the application to manage the life cycle of messages sent in this fashion. Throttling is enabled if the throttle value is greater than “0”. Throttling occurs on the default threads and the request will be sent from there. Setting the throttle rate to “0” will disable throttling and the request will be sent from the thread creating the subscription.

class SendComplete : public MamaSendCompleteCallback 
{ 
public:
    void onSendComplete (MamaPublisher& publisher, 
                         MamaMsg*       msg,
                         MamaStatus&    status, 
                         void*          closure) 
    { 
        // Check the status of the call. 
        // Destroy the message or reuse. 
    }
}

MamaSendCompleteCallback* cb = new MamaSendCompleteCallback; 

static void MAMACALLTYPE sendCompleteCb (mamaPublisher publisher,
                                         mamaMsg       msg,
                                         mama_status   status,
                                         void*         closure)
{
      // Check the status of the call. 
      // Destroy the message or reuse. 
}

mamaPublisher_sendWithThrottle (publisher,
                                msg,
                                sendCompleteCb,
                                NULL /*No need to use closure*/);

//To throttle sending of message
publisher->sendWithThrottle (msg, cb, closure);

class SendComplete : MamaSendCompleteCallback
{
    public void onSendComplete (MamaPublisher publisher,
                                MamaMsg msg,
                                MamaStatus.mamaStatus status,
                                object closure)
    {
        // Check the status of the call.
        // Destroy the message or reuse.
    }
}
MamaSendCompleteCallback cb = new MamaSendCompleteCallback ();
publisher.sendWithThrottle (msg, cb, closure);

class SendComplete : MamaSendCompleteCallback
{
    public void onSendComplete (MamaPublisher publisher,
                                MamaMsg msg,
                                MamaStatus.mamaStatus status,
                                object closure)
    {
        // Check the status of the call.
        // Destroy the message or reuse.
    }
}
MamaSendCompleteCallback cb = new MamaSendCompleteCallback ();
publisher.sendWithThrottle (msg, cb, closure);

Request/Response

Request/response communication involves a requester of data issuing a request on a topic to responder (s) of data for that topic on a particular transport. A requester can receive multiple responses to a request. The response is always sent over unicast. Request/response communication in OpenMAMA is achieved through the use of MamaPublisher and MamaInbox objects.

To make a request in OpenMAMA, a message is sent from a publisher and is associated with an instance of an inbox. The inbox is the destination recipient for any responses to the issued request. A callback is registered with the inbox upon creation and is invoked whenever any responses to the request are received by the API. Inbox requests can be throttled in the same way as non point-to-point messages and are affected by the same threading conditions.

Request messages arrive to the onMsg() callback for basic subscriptions on particular topics. A message is identified as being a request via a call to the isFromInbox() method. The sendReplyToInbox() method is used to actually send the response message to the request. Both the request message and the new response message are passed to the method when it is invoked.

Advanced Publishing

For advanced, or data quality publishing, the DQPublisherManager is the central class for advanced publishing. It is responsible for the namespace subscription, listening for subscription requests, and handles the DQPublishers, created to respond to those requests. There are also mechanisms for handling refresh messages and synch requests.

Creating a publisher manager requires a transport, a queue, a symbol namespace, the necessary callbacks and an optional root. The root is an identifier specific to the platform. The default is _MD, which is what the client applications subscribe to. For publishing a dictionary the default is _MDDD.

auto pubManager = new MamaDQPublisherManager();
pubManager->create (transport,
                    queue,
                    callbacks,
                    sourcename);

mamaDQPublisherManager_allocate(&pubManager); 
mamaDQPublisherManager_create (pubManager,
                               transport, 
                               queue,
                               callbacks, 
                               sourcename, 
                               root, 
                               NULL);

MamaDQPublisherManager pubManager = new MamaDQPublisherManager();
pubManager.create (transport, 
                   queue,
                   callbacks,
                   sourcename);

// Not currently supported on C#

Upon creation, the publisher manager creates a subscription to listen for requests on the given namespace. When a new request is received by the publisher manager it calls an onNewRequest callback. The callback supplies the symbol from the request as well as the type and request type. The message, containing the IP address of the requesting client machine, is also supplied.

class SubscriptionHandler : public MamaDQPublisherManagerCallback
{
    ...
    void onNewRequest (MamaDQPublisherManager* publisherManager,
                       const char*             symbol,
                       short                   subType,
                       short                   msgType,
                       MamaMsg&                msg)
    {
        // Handle the new request
    }
    ...
}

static void MAMACALLTYPE
onNewRequestCb (mamaDQPublisherManager pubManager, 
                const char*            symbol, 
                short                  subType, 
                short                  msgType,
                mamaMsg                msg)
{
    // Handle the new request
}

private static class SubscriptionHandler extends MamaDQPublisherManagerCallback
{
    ...
    public void onNewRequest (MamaDQPublisherManager publisherManager,
                              String symbol,
                              short subType,
                              short msgType,
                              MamaMsg msg)
    {
        // Handle the new request
    }
    ...
}

// Not currently supported on C#

When the request has been received, and if the publishing application can supply data for the requested symbol, then the publishing application should create a publisher and add it to the manager.

pubManager->createPublisher (symbol, closureData);

mamaDQPublisher dqPublisher;
mamaDQPublisherManager_createPublisher (pubManager, symbol, closureData, &dqPublisher);

pubManager.createPublisher(symbol, closureData);

// Not currently supported on C#

If the symbol should not be published, then the request should be ignored and the subscription will timeout on the client side.

A caching structure may also be added at this point. Neither the publisher or the publishing manager interact with the caching structure directly, and it is up to the client application to use a structure which best suits that application. The example applications can use either a simple message or a full field cache.

When creating the publisher, it can be configured in a number of ways. Setting the initial seqnum to “0” means that seqnum is not sent and this effectively turns off recapping. Initials and updates are handled as normal, but if a gap is detected the client will be unable to recover. By default a seqnum is added to every message. Sender ID is also added to every message, unless it is set to “0” explicitly. The msgstatus, also present in every message, is set to whatever the status on the publisher is at the time of send. Default status is MAMA_STATUS_OK, but this can be changed to STALE or MAYBE_STALE as required.

pubManager->setStatus (status); 
pubManager->setSenderId (senderid); 
pubManager->setSeqNum (num); 

mamaDQPublisherManager_setStatus (pubManager, status); 
mamaDQPublisherManager_setSenderId (pubManager, senderid); 
mamaDQPublisherManager_setSeqNum (pubManager, num);

pubManager.setStatus (status); 
pubManager.setSenderId (senderid); 
pubManager.setSeqNum (num); 

// Not currently supported on C#

The first message published in response to a new request should be of type INITIAL or RECAP. This is a full image of the caching structure and is sent using the send reply method.

initialMsg->updateU8 (NULL, MamaFieldMsgType.mFid, MAMA_MSG_TYPE_INITIAL); 
symbolPub->sendReply (msg, initialMsg);

mamaMsg_updateU8 (initialMsg, NULL, MamaFieldMsgType.mFid, MAMA_MSG_TYPE_INITIAL); 
mamaDQPublisher_sendReply(symbolPub, msg, initialMsg);

initialMsg.updateU8 (NULL, MamaFieldMsgType.mFid, MAMA_MSG_TYPE_INITIAL); 
symbolPub.sendReply (msg, initialMsg);

// Not currently supported on C#

Once the first message is sent, the client application listens for updates published from that publisher. The recommended approach is for updates to be applied to the cache then the delta to be sent.

When a request for a symbol that is already being published and is familiar to the publisher manager is made, the OnRequest callback is called. As well as the parameters in the OnNewRequest callback, you are also given pointers to the publisher currently publishing this symbol and the closure structure for that symbol.

class SubscriptionHandler : public MamaDQPublisherManagerCallback
{
    ...
    void onRequest (MamaDQPublisherManager*  publisherManager,
                    const MamaPublishTopic&  publishTopicInfo,
                    short                    subType,
                    short                    msgType,
                    MamaMsg&                 msg)
    {
        // Handle the request
    }
    ...
}

static void MAMACALLTYPE
onRequestCb (mamaDQPublisherManager pubManager, 
             mamaPublishTopic*      publishTopicInfo, 
             short                  subType, 
             short                  msgType,
             mamaMsg                msg)
{
    // Handle the request
}

private static class SubscriptionHandler extends MamaDQPublisherManagerCallback
{
    ...
    public void onRequest (MamaDQPublisherManager                  publisherManager,
                           MamaDQPublisherManager.MamaPublishTopic mamaPublishTopic,
                           short                                   subType,
                           short                                   msgType,
                           MamaMsg                                 msg)
    {
        // Handle the request
    }
    ...
}

// Not currently supported on C#

If the publishing application wishes to publish to that client, then it must first send an initial. However, as the cache is being used to publish updates, it is important to make sure that you get a full image in a thread-safe way. The difference between OnNewRequest and OnRequest is that OnRequest has a publisher created and data, whereas OnNewRequest does not, so your data initialization should occur in OnNewRequest. Once the initial is published, the client processes the updates as normal.

Recap requests are passed up in the same way as requests for symbols already being published, through the OnRequest callback. This is because they are handled in the same way. The only difference between the two is that you set the type on the image message published as RECAP rather than INITIAL. You may also want to send the recap as multicast rather than as a reply.

Stop publishing (Refresh Mechanism)

To be in accordance with the normal publishing rules within the platform, publishers should cease to publish one hour after the last request or refresh. Refresh messages are sent by client applications approximately every 55 minutes, and the onRefresh callback will be fired.

class SubscriptionHandler : public MamaDQPublisherManagerCallback
{
    ...
    void onRefresh (MamaDQPublisherManager*  publisherManager,
                    const MamaPublishTopic&  publishTopicInfo,
                    short                    subType,
                    short                    msgType,
                    MamaMsg&                 msg)
    {
        // Handle the event
    }
    ...
}

static void MAMACALLTYPE
onRefreshCb (mamaDQPublisherManager pubManager, 
             mamaPublishTopic*      publishTopicInfo, 
             short                  subType, 
             short                  msgType,
             mamaMsg                msg)
{
    // Handle the event
}

private static class SubscriptionHandler extends MamaDQPublisherManagerCallback
{
    ...
    public void onRefresh (MamaDQPublisherManager                  publisherManager,
                           MamaDQPublisherManager.MamaPublishTopic mamaPublishTopic,
                           short                                   subType,
                           short                                   msgType,
                           MamaMsg                                 msg)
    {
        // Handle the event
    }
    ...
}

// Not currently supported on C#

There is an acknowledgement sent to all clients for that symbol that refreshing has occurred, which is handled internally. The publishing application should record when the last refresh or request for a symbol was received, and if it was more than one hour ago it may stop publishing that symbol and remove it from the publisher manager.

If the symbol is requested again, an OnNewRequest callback is generated.

If a refresh is received for a symbol that is not currently being published, a OnNewRequest callback is received rather than an OnRefresh. This is useful when bringing up a secondary source while not putting extra strain on the client applications to respond to synchronization request messages.

Synchronization requests can be sent by the publisher manager during start-up to determine which clients are currently subscribed to it. This is generally done when a source has been restarted. A request is sent by the MamaDQPublisher Manager and all clients subscribed respond with a list of the symbols they are interested in. The synchronization request message contains some values to help separate the reply messages if there are a lot of subscriptions across multiple clients.

The reply is parsed by the publisher manager and will generate a number of OnNewRequest and OnRequest callbacks.

It is important that synchronization requests are only sent when the publisher is in a position to handle all the requests generated, and that they are sent after the transport is fully connected. For example, on a TCP connection you would wait for the transport connected callback.

The OpenMAMA API provides a fault-tolerant module which can be used to heartbeat between two applications and tell which is primary and which is secondary (see Section 12.2: Data quality and fault tolerant takeovers). It is important that when attempting a fault-tolerant takeover that at a minimum you recap everything that has ticked since the last heartbeat. This is due to the fact that OpenMAMA clients will not request a recap when there is a change in seqnum that coincides with a change in sender id. It is the responsibility of the publisher to maintain data quality through a takeover event.

Value Add

Timers

OpenMAMA supports user events triggered by recurring timers. These are represented by the MamaTimer object. An event queue must be specified when a timer is created.

A timer is created with a specified interval in seconds, using a double-precision floating point number to give the resolution in fractions of a second if required. The timer callback is invoked repeatedly at an interval no shorter than that specified. A number of factors can cause the timer interval to be inexact or the callback not to be immediately invoked on the firing of the timer.

The timer itself can be inaccurate due to the frequency of the OS interrupts. For instance, on a 100 Hz OS (Linux 2.4), if a 10 milliseconds timer is created one millisecond after the last interrupt fired it will take 19 milliseconds for the timer to fire. This is because only 9 milliseconds will have elapsed when the next scheduled interrupt occurs. Essentially, the best timer resolution possible here is 10 milliseconds and worst 19.9999 milliseconds.

The precision of timers is determined by the implementation in the underlying messaging platform and the interrupt frequency of the operating system.

For example, the Linux 2.4 kernel with an interrupt frequency of 100 Hz is capable of 10 milliseconds resolution at best. A Linux 2.6 (up to 2.6.12) kernel with a 1000 Hz interrupt frequency can provide timers with a best resolution of 1 millisecond. (The default interrupt frequency on Linux 2.6.13 and up is 250 Hz but is now configurable on i386 architectures).

The invoking of a callback in response to a timer firing can depend on the activity on the event queue which was specified on timer creation.

A timer can be destroyed from within a timer callback or any other callback on the queue. This function must be called from the same thread dispatching on the associated event queue unless both the default queue and dispatch queue are not actively dispatching.

Creating a Timer

The following example show the creation of a recurring timer which fires every 500 milliseconds. In each case the actionCallback function/object receives a callback once the interval has elapsed.

class TimerCallback : public MamaTimerCallback 
{
public:
    TimerCallback () {} 
    virtual ~TimerCallback () {} 
    virtual void onTimer (MamaTimer* timer) 
    { 
        MyClosureType* myClosure = dynamic_cast<MyClosureType*>(timer->getClosure () 
        // Perform recurring task. 
    }
}
....
MamaTimer* timer = new MamaTimer;
TimerCallback* timerCallback = new TimerCallback;
timer->create (queue,
               timerCallback,
               0.5,
               closure);

static void actionCallback (mamaTimer timer,
                            void*     closure)
{
    myClosureType* myClosure = (myClosureType*)closure;
    // Perform recurring task.
}

mamaTimer_create (&timer,
                  queue,
                  actionCallback,
                  0.5,
                  myClosure);

class TimerCallback implements MamaTimerCallback
{
    private Object myContextData = null;
    public TimerCallback (Object contextData)
    {
        myContextData = contextData;
    }
    public void onTimer (MamaTimer timer)
    {
        //perform recurring task
    }
}
....
MamaTimer timer = new MamaTimer ();
TimerCallback timerCallback = new TimerCallback ();
timer.create (queue, timerCallback, 0.5);

class TimerCallback : MamaTimerCallback
{
    private object myContextData = null;
    public TimerCallback (object contextData)
    {
        myContextData = contextData;
    }
    public void onTimer (MamaTimer timer, object closure)
    {
        //perform recurring task
    }
}
....
MamaTimer timer = new MamaTimer ();
TimerCallback timerCallback = new TimerCallback();
timer.create (queue, timerCallback, 0.5);

IO

The OpenMAMA API provides an abstract mechanism by which a client registers interest for various events on file descriptors. A callback, provided to the API, is invoked whenever an event, of the type specified when registering interest, occurs. The MamaIo API facilitates asynchronous IO operations.

The following event types are supported in the API:

• MAMA_IO_READ
• MAMA_IO_WRITE
• MAMA_IO_CONNECT
• MAMA_IO_ACCEPT
• MAMA_IO_CLOSE
• MAMA_IO_ERROR
• MAMA_IO_EXCEPT

Registering for IO Events

Not all underlying messaging middlewares support all of the event types provided by the OpenMAMA API. In the case of a particular event type not being supported, the call to create() fails with a return code of MAMA_STATUS_UNSUPPORTED_IO_TYPE.

class IoCallback : public MamaIoCallback
{
    virtual void onIo (MamaIo* io, mamaIoType ioType)
    {
        // The inboundFileDescriptor is now available for reading.
    }
}
MamaIo mamaIo = new MamaIo;
IoCallback ioCallback = new IoCallback;
int inboundFileDescriptor = 0;//File descriptor for socket we want to read from
....
// Get the file descriptor for an inbound accept on a socket
....
mamaIo->create (queue, ioCallback, inboundFileDescriptor, MAMA_IO_READ, "My Closure");

static void ioCallback (mamaIo io,
                        mamaIoType ioType,
                        void* closure)
{
    // The inboundFileDescriptor is now available for reading.
}
mamaIo readHandler = NULL;

// File descriptor for socket we want to read from
int inboundFileDescriptor = 0;
....
// Get the file descriptor for an inbound accept on a socket
....
mamaIo_create (&readHandler,
               queue,
               inboundFileDescriptor,
               ioCallback,
               MAMA_IO_READ,
               "My Closure");

// This API is not supported on Java

class IoCallback : MamaIoCallback
{
    public void onIo (MamaIo io, mamaIoType ioType)
    {
        // The inboundFileDescriptor is now available for reading.
    }
}
MamaIo mamaIo = new MamaIo ();
IoCallback ioCallback = new IoCallback ();
int inboundFileDescriptor = 0;//File descriptor for socket we want to read from
....
//Get the file descriptor for an inbound accept on a socket
....
mamaIo.create (queue,
               ioCallback,
               inboundFileDescriptor,
               mamaIoType.MAMA_IO_READ,
               "My Closure");

User Events

User events can be added to OpenMAMA queues irrespective of the middleware using MamaQueue’s
enqueueEvent(). This allows user code to be executed on any of the dispatching threads.

MamaQueueEventCallback’s onEvent() method will be invoked whenever the event fires. The callback event is added to the back of the queue, so if there are a large number of events on the queue it may be some time before the callback is invoked.

Logging

To aid with debugging, the OpenMAMA API provides various levels of verbose logging to give developers more detail during event processing within the API.By default, logging is disabled within the API and needs to be manually enabled along with specifying the level at which to log information.

Logging Via Properties

There are several values that can be set in mama.properties to control logging. They are listed in Table 14: Logging Properties. Logs can be set to roll over. When set to roll over, the following happens:

1. The log file is allowed to grow until it reaches the maximum size.
2. When it reaches the maximum size, the logfile has .1 appended to its name and a new log file is begun.
3. If a .1 file already exists, this older file is renamed to have a .2 instead of .1, and so on.
4. If there is a maximum number of log files specified, then the total number of log files will never be more than that. The oldest file will then be deleted when the log file rolls over

If using an UNBOUNDED policy, the behaviour is undefined if the log file reaches the maximum size allowed by the file system.

There are several methods used to control logging. They are listed in the following table.

Conflation

Conflation from the OpenMAMA Client Perspective

Conflation is the process employed by an advanced publisher to merge messages on the write queue to reduce memory use, to reduce bandwidth, and to deal with slow consumers. Conflation is available for the NYSE Technologies Data Fabric.

The advanced publisher installs and un-installs a conflater for individual clients as and when they require conflation. There are two circumstances that require conflation:

1. If the number of messages increases above a high water mark (the advanced publisher write queue is becoming too large).
2. The client requests that conflation is turned on. A client may monitor its own queue sizes to determine when it requires conflated data.

From a OpenMAMA client perspective, the most important aspect is the method by which a client requests conflation. There is a OpenMAMA function available for this.

transport.requestConflation();

There are two reasons why the conflater may no longer be needed:

1. When the number of messages drops below a low water mark.
2. When the client requests that conflation is no longer needed. There is an OpenMAMA function available to request the end of conflation. Example 48: Ending conflation

transport.requestEndConflation();

Although clients can request conflation or request that the advanced publisher stops conflating messages at any time, the advanced publisher may not be able to honor the request. For example, some advanced publishers may not be configured for conflation while others may be unable to end conflation at a client’s request because the write queues are too large.

When a client receives a conflated message it contains three additional fields: wConflateCount, wConflateTradeCount, and wConflateQuoteCount. These represent the total number of messages conflated into a single message, the number of trades and the number of quotes respectively. wConflateCount == wConflateTradeCount + wConflateQuoteCount. OpenMAMA and OpenMAMDA use these fields internally to perform sequence number checking and maintain data quality.

Statistics

OpenMAMA provides statistics to monitor the OpenMAMA client. When enabled, OpenMAMA will log, via OpenMAMA logging, and/or publish details of the numbers of subscription requests received, number of initial messages received, queue size and various other statistics. 29West LBM also provides various low-level statistics on a per context basis which are exposed via this functionality. Statistics are generated and logged and/or published at a configurable interval. When OpenMAMA statistics are logged, the value of the statistic for the interval is logged, along with the maximum and running total for that statistic since statistics logging was enabled. When OpenMAMA statistics are published, the interval value for each statistic is published along with a timestamp indicating the time the message was generated and information identifying the client such as the IP Address, user name, and OpenMAMA application name. When used in conjunction with the Stats Logger, customers can produce DSV files containing client monitoring information.

Note that the size of the published message can be large. When publishing over 29West LBM, it is likely that such a message will be too large to publish using immediate messaging. Depending on the 29West LBM version being used, this will result in either the message not being published at all, or the message being published in a truncated form and thus missing some or all of the stats vector. For this reason, we recommend disabling the use of immediate messaging for publishing over 29West LBM using the property mama.lbm.transport.<transport name>.use_im_for_publishing=false.

Statistics are provided at various levels: globally (all transports and queues across the application), per- transport, and per-queue. The level at which to generate statistics is configurable, and any combination of global, transport and queue statistics can be used. 29West LBM statistics are enabled separately but are considered transport-level statistics.

Note that when logged to the OpenMAMA log, only statistics with a value greater than “0” will appear in the log.

General OpenMAMA Statistics

Statistics Logging Configuration

The following parameters controls statistics logging. They can be set in mama.properties.

Published Message Format

Stats reports can be published by enabling one or more of the publishing parameters on the middleware, specified by mama.statslogging.middleware, and transport, specified by mama.statslogging.transport. These can be published to the Stats Logger product to automatically generate DSV files containing the stats.

These stats reports are simply OpenMAMA messages published on the topic STATS_TOPIC, and so it is possible to write a OpenMAMA application to subscribe to stats messages and manually process them. The format of the published messages is described in the following table.

The following table describes the messages included in MamaStatEvents. All the fields described in the table are available from the MamaStatFields header. Names and FIDs for individual fields can my accessed using FIELD.mName and FIELD.mFid.

Example Programs

This section gives a brief explanation of what each of the OpenMAMA example programs is designed to show. Each example is provided as a prebuilt binary and as source code within the release. Each example accepts a list of command line options. These can be examined by passing -h on the command line, or viewing the source code file.

Note The examples should be available across all languages, however, some may have a slight name change, such as mamalistenc, mamalistencpp, mamalistencs, and mamalistenjava. The operation and functionality across the languages should be consistent.

Performance Programming

OpenMAMA and OpenMAMDA are commonly used within applications where performance is important. OpenMAMA itself is very lightweight and an application that does significant work will use much more CPU cycles that the API itself. This section describes a number of techniques and tips that can be used to optimize the use of OpenMAMA, OpenMAMDA and applications built on top of them.

Monitoring Performance

There are a number of ways to monitor the performance of a OpenMAMA or OpenMAMDA application:

• A good indicator of performance is the queue depth: if the queue is growing then it is a sign that the application is not able to keep up with the amount of incoming data. This adds latency. From within the API, either the depth of queues can be queried or watermarks can be set for a particular queue depth value.
• MamaStats provides a lot of useful information including the queue depth, message rates, number of recaps as well as middleware-specific statistics.
• Recap requests, shown via the MamaStats output or thought the onRecapRequest() callback, are also an indicator that the application is not able to keep up with the data. Recaps are not a good indicator of performance because, by the time a recap is requested, permanent data loss has already occurred.
• The application’s CPU use should be monitored. If any of the cores used by the application approach or reach 100% use then issues such as queue growth and data loss will occur.

Storing Per-Symbol State

It is usual to need to store application state on a per symbol basis e.g. calculated values, pointer/ references to other application objects etc.

Note It is very inefficient to do this by obtaining the symbol from each received message and then using a map to retrieve the state for that symbol.

OpenMAMA and OpenMAMDA provide an alternative mechanism for storing state though closures. A closure is a user-defined void pointer and is associated with a particular symbol.

Closures can be set on a subscription, created and then retrieved using the accessors on the subscription objects. Single and group subscriptions can be used to receive data from many different symbols, therefore a separate “item closure” is available for the group members. Calling setItemClosure() in a group subscription callback will set the closure for that group member.

Another method available is to store the per-symbol state within the objects which implement the subscription callback interfaces e.g. the abstract base class MamaSubscriptionCallback object oriented implementations. To do this, a separate instance of the callback object should be created for each subscription and its state should be stored in its member variables.

Message Access

If accessing all, or most, of the data within a message, then iteration is generally faster than direct field access. There are two methods: user driven and callback driven. User driven has been shown to be faster.

If accessing only a few fields within a message, then direct access to those particular fields will generally be faster. The decision of which method to choose will depend on two factors: the particular application use case and the deployment, which will determine the number of fields and where those fields are within a message, so experimentation is encouraged.

Memory Allocation

Memory allocation and deallocation are major sources of performance issues. It is best to allocate all required memory at start up, and avoid allocating memory while processing data, particularly in on a per- message basis. There are various ways to do this, which include: reusing the same specific objects from one callback to another; or, if the required lifetime of data stored within an object is longer than the scope of a callback, then a pool of reusable objects can be used.

Threading

The recommended way to scale a OpenMAMA or OpenMAMDA application across using multiple CPU cores is to use separate threads on mamaQueues, with one thread/queue per CPU core allocated to OpenMAMA. For example, on an eight core machine with a non-MAMA application having two “hot” threads, the OpenMAMA application having two “hot” threads but not directly using OpenMAMA, then three mamaQueues should be used. The definition of a “hot” thread is one that continuously uses a significant amount of CPU. The eighth core in the example above is left unassigned because it is recommended to reserve a core for the operating system and other threads that don’t use a significant amount of CPU.

Again, different applications will have behave differently so some experimentation may be required to find the optimal number of queues. The example applications supplied with OpenMAMA demonstrate an easy way to do this how by having the number of queues configurable number of queues at runtime using the MamaQueueGroup object.

An alternative method is to use a separate thread to do the message processing instead of processing the data in the callbacks. This is not generally recommended as it just shifts the bottleneck from one thread to another, and causes additional overhead by having to copy the message or fields from the message and place onto another queue.

Running Multiple Instances of OpenMAMA

Only one instance of OpenMAMA is supported for each process. However, multiple OpenMAMA processes can be run successfully.

By default, when MAMA’s open function is called, the mama.properties file is obtained from the location specified by the WOMBAT_PATH environment variable. This results in all OpenMAMA processes being configured using the same file.

Running with a Single Properties File

This is the default approach taken when running multiple instances of the OpenMAMA example programs, such as mamalistenc.

OpenMAMA still runs, but with the following restrictions:

1. Only a single log file can be specified.

   The mama.logging.file.name entry in mama.properties can be used to specify the log file. All OpenMAMA processes attempt to open and write to the log file at the same time. To prevent this occuring:

• Do not specify a log file. If a log file is not specified OpenMAMA writes log entries to stderr, the output can be directed on a per-process basis.
• Change the location of the log file in code using the MAMA’s logToFile function.
• Install a log callback function in code to catch all logging events using the mama_setLogCallback function.

1. Publishing features may not work properly.

   OpenMAMA provides several features that create transports as part of their initialization. Transport names and settings are read from mama.properties when MAMA’s open function is called.

   This is particularly a problem publishing data when using a point to point middleware such as WMW TCP. It results in an error being returned from MAMA’s open function as the publishing port is blocked by a prior OpenMAMA instance.

   The following features publish data on the transport. If any of these are to be used at the same time across multiple OpenMAMA instances, then consideration should be given to maintaining multiple mama.properties files or configuration of the API in code.

   • Stats publishing
   • Usage logging
   • WAM

The following features create subscriptions to request data. In these cases a multi-cast middleware could be used that allows multiple connections to be open at the same time.

• Property server connection
• Template server connection

Managing Multiple Properties Files

Due to the restrictions highlighted in the Running with a Single Properties File section, it may be desirable to configure each process independently by supporting multiple properties files. OpenMAMA provides two methods of doing this:

• As explained in the previous section, the default behaviour of MAMA’s open can be overridden using MAMA’s openWithProperties() function. This function takes a file name and location of the properties file.
• Alternatively, each process can define a different location in the WOMBAT_PATH environment

Configuration Reference

OpenMAMA properties that apply to all middlewares are listed in the following table.

OpenMAMA Status Codes

This section lists the status and error codes. These are defined in mama/status.h.

Glossary