Persistent messages guarantee message delivery in case of broker failure. The broker stores the message in a persistent store until all intended consumers acknowledge they have consumed the message.

Broker processing of persistent messages is slower than for nonpersistent messages for the following reasons:

For both queues and topics with durable subscribers, performance was approximately 40% faster for nonpersistent messages. We obtained these results using 10k-sized messages and AUTO_ACKNOWLEDGE mode.

A transaction is a guarantee that all messages produced in a transacted session and all messages consumed in a transacted session will be either processed or not processed (rolled back) as a unit.

Message Queue supports both local and distributed transactions.

A message produced or acknowledged in a transacted session is slower than in a nontransacted session for the following reasons:

One mechanism for ensuring the reliability of JMS message delivery is for a client to acknowledge consumption of messages delivered to it by the Message Queue broker.

If a session is closed without the client acknowledging the message or if the broker fails before the acknowledgment is processed, the broker redelivers that message, setting a JMSRedelivered flag.

For a nontransacted session, the client can choose one of three acknowledgment modes, each of which has its own performance characteristics:

(Using CLIENT_ACKNOWLEDGE mode is similar to using transactions, except there is no guarantee that all acknowledgments will be processed together if a provider fails during processing.)

Acknowledgment mode affects performance for the following reasons:

Subscribers to a topic destination fall into two categories, those with durable and nondurable subscriptions.

Durable subscriptions provide increased reliability but slower throughput, for the following reasons:

We compared performance for durable and nondurable subscribers in two cases: persistent and nonpersistent 10k-sized messages. Both cases use AUTO_ACKNOWLEDGE acknowledgment mode. We found an effect on performance only in the case of persistent messages which slowed durables by about 30%

Application developers often want to target sets of messages to particular consumers. They can do so either by targeting each set of messages to a unique physical destination or by using a single physical destination and registering one or more selectors for each consumer.

A selector is a string requesting that only messages with property values that match the string are delivered to a particular consumer. For example, the selector NumberOfOrders>1 delivers only the messages with a NumberOfOrders property value of 2 or more.

Creating consumers with selectors lowers performance (as compared to using multiple physical destinations) because additional processing is required to handle each message. When a selector is used, it must be parsed so that it can be matched against future messages. Additionally, the message properties of each message must be retrieved and compared against the selector as each message is routed. However, using selectors provides more flexibility in a messaging application.

Message size affects performance because more data must be passed from producing client to broker and from broker to consuming client, and because for persistent messages a larger message must be stored.

However, by batching smaller messages into a single message, the routing and processing of individual messages can be minimized, providing an overall performance gain. In this case, information about the state of individual messages is lost.

In our tests, which compared throughput in kilobytes per second for 1k, 10k, and 100k-sized messages to a queue destination and AUTO_ACKNOWLEDGE acknowledgment mode, we found that nonpersistent messaging was about 50% faster for 1k messages, about 20% faster for 10k messages, and about 5% faster for 100k messages. The size of the message affected performance significantly for both persistent and nonpersistent messages. 100k messages are about 10 times faster than 10k, and 10k are about 5 times faster than 1k.

JMS supports five message body types, shown below roughly in the order of complexity:

While, in general, the message type is dictated by the needs of an application, the more complicated types (MapMessage and ObjectMessage) carry a performance cost: the expense of serializing and deserializing the data. The performance cost depends on how simple or how complicated the data is.

For both the Message Queue broker and client applications, CPU processing speed and available memory are primary determinants of message service performance. Many software limitations can be eliminated by increasing processing power, while adding memory can increase both processing speed and capacity. However, it is generally expensive to overcome bottlenecks simply by upgrading your hardware.

Because of the efficiencies of different operating systems, performance can vary, even assuming the same hardware platform. For example, the thread model employed by the operating system can have an important effect on the number of concurrent connections a broker can support. In general, all hardware being equal, Solaris is generally faster than Linux, which is generally faster than Windows.

The broker is a Java process that runs in and is supported by the host JVM. As a result, JVM processing is an important determinant of how fast and efficiently a broker can route and deliver messages.

In particular, the JVM’s management of memory resources can be critical. Sufficient memory has to be allocated to the JVM to accommodate increasing memory loads. In addition, the JVM periodically reclaims unused memory, and this memory reclamation can delay message processing. The larger the JVM memory heap, the longer the potential delay that might be experienced during memory reclamation.

The number and speed of connections between client and broker can affect the number of messages that a message service can handle as well as the speed of message delivery.

Broker Connection Limits

All access to the broker is by way of connections. Any limit on the number of concurrent connections can affect the number of producing or consuming clients that can concurrently use the broker.

The number of connections to a broker is generally limited by the number of threads available. Message Queue can be configured to support either a dedicated thread model or a shared thread model (see Thread Pool Management).

The dedicated thread model is very fast because each connection has dedicated threads, however the number of connections is limited by the number of threads available (one input thread and one output thread for each connection). The shared thread model places no limit on the number of connections, however there is significant overhead and throughput delays in sharing threads among a number of connections, especially when those connections are busy.

Transport Protocols

Message Queue software allows clients to communicate with the broker using various low-level transport protocols. Message Queue supports the connection services (and corresponding protocols) described in Configuring Connection Services.

The choice of protocols is based on application requirements (encrypted, accessible through a firewall), but the choice affects overall performance.

Our tests compared throughput for TCP and SSL for two cases: a high-reliability scenario (1k persistent messages sent to topic destinations with durable subscriptions and using AUTO_ACKNOWLEDGE acknowledgment mode) and a high-performance scenario (1k nonpersistent messages sent to topic destinations without durable subscriptions and using DUPS_OK_ACKNOWLEDGE acknowledgment mode).

In general we found that protocol has less effect in the high-reliability case. This is probably because the persistence overhead required in the high-reliability case is a more important factor in limiting throughput than the protocol speed. Additionally:

A Message Queue message service can be implemented as a single broker or as a cluster consisting of multiple interconnected broker instances.

As the number of clients connected to a broker increases, and as the number of messages being delivered increases, a broker will eventually exceed resource limitations such as file descriptor, thread, and memory limits. One way to accommodate increasing loads is to add more broker instances to a Message Queue message service, distributing client connections and message routing and delivery across multiple brokers.

In general, this scaling works best if clients are evenly distributed across the cluster, especially message producing clients. Because of the overhead involved in delivering messages between the brokers in a cluster, clusters with limited numbers of connections or limited message delivery rates, might exhibit lower performance than a single broker.

You might also use a broker cluster to optimize network bandwidth. For example, you might want to use slower, long distance network links between a set of remote brokers within a cluster, while using higher speed links for connecting clients to their respective broker instances.

For more information on clusters, see Configuring and Managing Broker Clusters

The message throughput that a broker might be required to handle is a function of the use patterns of the messaging applications the broker supports. However, the broker is limited in resources: memory, CPU cycles, and so forth. As a result, it would be possible for a broker to become overwhelmed to the point where it becomes unresponsive or unstable.

The Message Queue message broker has mechanisms built in for managing memory resources and preventing the broker from running out of memory. These mechanisms include configurable limits on the number of messages or message bytes that can be held by a broker or its individual physical destinations, and a set of behaviors that can be instituted when physical destination limits are reached.

With careful monitoring and tuning, these configurable mechanisms can be used to balance the inflow and outflow of messages so that system overload cannot occur. While these mechanisms consume overhead and can limit message throughput, they nevertheless maintain operational integrity.

Message Queue supports both file-based and JDBC-based persistence modules. File-based persistence uses individual files to store persistent data. JDBC-based persistence uses a Java Database Connectivity (JDBC) interface and requires a JDBC-compliant data store. File-based persistence is generally faster than JDBC-based; however, some users prefer the redundancy and administrative control provided by a JDBC-compliant store.

In the case of file-based persistence, you can maximize reliability by specifying that persistence operations synchronize the in-memory state with the data store. This helps eliminate data loss due to system crashes, but at the expense of performance.

The Message Queue client runtime provides client applications with an interface to the Message Queue message service. It supports all the operations needed for clients to send messages to physical destinations and to receive messages from such destinations. The client runtime is configurable (by setting connection factory attribute values), allowing you to control aspects of its behavior, such as connection flow metering, consumer flow limits, and connection flow limits, that can improve performance and message throughput. See Client Runtime Message Flow Adjustments for more information on these features and the attributes used to configure them.

See your system documentation for tuning your operating system.

By default, the broker uses a JVM heap size of 192MB. This is often too small for significant message loads and should be increased.

When the broker gets close to exhausting the JVM heap space used by Java objects, it uses various techniques such as flow control and message swapping to free memory. Under extreme circumstances it even closes client connections in order to free the memory and reduce the message inflow. Hence it is desirable to set the maximum JVM heap space high enough to avoid such circumstances.

However, if the maximum Java heap space is set too high, in relation to system physical memory, the broker can continue to grow the Java heap space until the entire system runs out of memory. This can result in diminished performance, unpredictable broker crashes, and/or affect the behavior of other applications and services running on the system. In general, you need to allow enough physical memory for the operating system and other applications to run on the machine.

In general it is a good idea to evaluate the normal and peak system memory footprints, and configure the Java heap size so that it is large enough to provide good performance, but not so large as to risk system memory problems.

To change the minimum and maximum heap size for the broker, use the -vmargs command line option when starting the broker. For example:

This command will set the starting Java heap size to 256MB and the maximum Java heap size to 1GB.

In any case, verify settings by checking the broker’s log file or using the imqcmd metrics bkr -m cxn command.

Once a protocol that meets application needs has been chosen, additional tuning (based on the selected protocol) might improve performance.

A protocol’s performance can be modified using the following three broker properties:

For TCP and SSL protocols, these properties affect the speed of message delivery between client and broker. For HTTP and HTTPS protocols, these properties affect the speed of message delivery between the Message Queue tunnel servlet (running on a Web server) and the broker. For HTTP/ HTTPS protocols there are additional properties that can affect performance (see HTTP/HTTPS Tuning).

The protocol tuning properties are described in the following sections.

nodelay

The nodelay property affects Nagle’s algorithm (the value of the TCP_NODELAY socket-level option on TCP/IP) for the given protocol. Nagle’s algorithm is used to improve TCP performance on systems using slow connections such as wide-area networks (WANs).

When the algorithm is used, TCP tries to prevent several small chunks of data from being sent to the remote system (by bundling the data in larger packets). If the data written to the socket does not fill the required buffer size, the protocol delays sending the packet until either the buffer is filled or a specific delay time has elapsed. Once the buffer is full or the timeout has occurred, the packet is sent.

For most messaging applications, performance is best if there is no delay in the sending of packets (Nagle’s algorithm is not enabled). This is because most interactions between client and broker are request/response interactions: the client sends a packet of data to the broker and waits for a response. For example, typical interactions include:

For these interactions, most packets are smaller than the buffer size. This means that if Nagle’s algorithm is used, the broker delays several milliseconds before sending a response to the consumer.

However, Nagle’s algorithm may improve performance in situations where connections are slow and broker responses are not required. This would be the case where a client sends a nonpersistent message or where a client acknowledgment is not confirmed by the broker (DUPS_OK_ACKNOWLEDGE session).

inbufsz/outbufsz

The inbufsz property sets the size of the buffer on the input stream reading data coming in from a socket. Similarly, outbufsz sets the buffer size of the output stream used by the broker to write data to the socket.

In general, both parameters should be set to values that are slightly larger than the average packet being received or sent. A good rule of thumb is to set these property values to the size of the average packet plus 1 kilobyte (rounded to the nearest kilobyte). For example, if the broker is receiving packets with a body size of 1 kilobyte, the overall size of the packet (message body plus header plus properties) is about 1200 bytes; an inbufsz of 2 kilobytes (2048 bytes) gives reasonable performance. Increasing inbufsz or outbufsz greater than that size may improve performance slightly, but increases the memory needed for each connection.

HTTP/HTTPS Tuning

In addition to the general properties discussed in the previous two sections, HTTP/HTTPS performance is limited by how fast a client can make HTTP requests to the Web server hosting the Message Queue tunnel servlet.

A Web server might need to be optimized to handle multiple requests on a single socket. With JDK version 1.4 and later, HTTP connections to a Web server are kept alive (the socket to the Web server remains open) to minimize resources used by the Web server when it processes multiple HTTP requests. If the performance of a client application using JDK version 1.4 is slower than the same application running with an earlier JDK release, you might need to tune the Web server keep-alive configuration parameters to improve performance.

In addition to such Web server tuning, you can also adjust how often a client polls the Web server. HTTP is a request-based protocol. This means that clients using an HTTP-based protocol periodically need to check the Web server to see if messages are waiting. The imq.httpjms.http.pullPeriod broker property (and the corresponding imq.httpsjms.https.pullPeriod property) specifies how often the Message Queue client runtime polls the Web server.

If the pullPeriod value is -1 (the default value), the client runtime polls the server as soon as the previous request returns, maximizing the performance of the individual client. As a result, each client connection monopolizes a request thread in the Web server, possibly straining Web server resources.

If the pullPeriod value is a positive number, the client runtime periodically sends requests to the Web server to see if there is pending data. In this case, the client does not monopolize a request thread in the Web server. Hence, if large numbers of clients are using the Web server, you might conserve Web server resources by setting the pullPeriod to a positive value.

For information on tuning the file-based persistent store, see Configuring a File-Based Data Store.

To configure physical destination limits, see the properties described in Physical Destination Properties.

If message producers tend to overrun message consumers, messages can accumulate in the broker. The broker contains a mechanism for throttling back producers and swapping messages out of active memory under low memory conditions, but it is wise to set a hard limit on the total number of messages (and message bytes) that the broker can hold.

Control these limits by setting the imq.system.max_count and the imq.system.max_size broker properties.

For example:

The defined value above means that the broker will only hold up to 5000 undelivered and/or unacknowledged messages. If additional messages are sent, they are rejected by the broker. If a message is persistent then the clinet runtime will throw an exception when the producer tries to send the message. If the message is non-persistent, the broker silently drops the message.

When an exception is thrown in sending a message, the client should process the exception by pausing for a moment and retrying the send again. (Note that the exception will never be due to the broker’s failure to receive a message; the exception is thrown by the client runtime before the message is sent to the broker.)

Messages sent and received by clients (payload messages), as well as Message Queue control messages, pass over the same client-broker connection. Delays in the delivery of control messages, such as broker acknowledgments, can result if control messages are held up by the delivery of payload messages. To prevent this type of congestion, Message Queue meters the flow of payload messages across a connection.

Payload messages are batched (as specified with the connection factory attribute imqConnectionFlowCount) so that only a set number are delivered. After the batch has been delivered, delivery of payload messages is suspended and only pending control messages are delivered. This cycle repeats, as additional batches of payload messages are delivered followed by pending control messages.

The value of imqConnectionFlowCount should be kept low if the client is doing operations that require many responses from the broker: for example, if the client is using CLIENT_ACKNOWLEDGE or AUTO_ACKNOWLEDGE mode, persistent messages, transactions, or queue browsers, or is adding or removing consumers. If, on the other hand, the client has only simple consumers on a connection using DUPS_OK_ACKNOWLEDGE mode, you can increase imqConnectionFlowCount without compromising performance.

There is a limit to the number of payload messages that the Message Queue client runtime can handle before encountering local resource limitations, such as memory. When this limit is approached, performance suffers. Hence, Message Queue lets you limit the number of messages per consumer (or messages per connection) that can be delivered over a connection and buffered in the client runtime, waiting to be consumed.

Consumer Flow Limits

When the number of payload messages delivered to the client runtime exceeds the value of imqConsumerFlowLimit for any consumer, message delivery for that consumer stops. It is resumed only when the number of unconsumed messages for that consumer drops below the value set with imqConsumerFlowThreshold.

The following example illustrates the use of these limits: consider the default settings for topic consumers:

When the consumer is created, the broker delivers an initial batch of 1000 messages (providing they exist) to this consumer without pausing. After sending 1000 messages, the broker stops delivery until the client runtime asks for more messages. The client runtime holds these messages until the application processes them. The client runtime then allows the application to consume at least 50% (imqConsumerFlowThreshold ) of the message buffer capacity (i.e. 500 messages) before asking the broker to send the next batch.

In the same situation, if the threshold were 10%, the client runtime would wait for the application to consume at least 900 messages before asking for the next batch.

The next batch size is calculated as follows:

So if imqConsumerFlowThreshold is 50%, the next batch size can fluctuate between 500 and 1000, depending on how fast the application can process the messages.

If the imqConsumerFlowThreshold is set too high (close to 100%), the broker will tend to send smaller batches, which can lower message throughput. If the value is set too low (close to 0%), the client may be able to finish processing the remaining buffered messages before the broker delivers the next set, again degrading message throughput. Generally speaking, unless you have specific performance or reliability concerns, you will not need to change the default value of imqConsumerFlowThreshold attribute.

The consumer-based flow controls (in particular, imqConsumerFlowLimit ) are the best way to manage memory in the client runtime. Generally, depending on the client application, you know the number of consumers you need to support on any connection, the size of the messages, and the total amount of memory that is available to the client runtime.

When the JMS resource adapter, jmsra, is used to consume messages in a GlassFish Server cluster, this behavior is defined using different properties, as described in About Shared Topic Subscriptions for Clustered Containers.

Connection Flow Limits

In the case of some client applications, however, the number of consumers may be indeterminate, depending on choices made by end users. In those cases, you can still manage memory using connection-level flow limits.

Connection-level flow controls limit the total number of messages buffered for all consumers on a connection. If this number exceeds the value of imqConnectionFlowLimit, delivery of messages through the connection stops until that total drops below the connection limit. (The imqConnectionFlowLimit attribute is enabled only if you set imqConnectionFlowLimitEnabled to true.)

The number of messages queued up in a session is a function of the number of message consumers using the session and the message load for each consumer. If a client is exhibiting delays in producing or consuming messages, you can normally improve performance by redesigning the application to distribute message producers and consumers among a larger number of sessions or to distribute sessions among a larger number of connections.