RabbitMQ PerfTest

RabbitMQ has a throughput testing tool, PerfTest, that is based on the Java client and can be configured to simulate basic workloads and more advanced workloads as well. PerfTest has extra tools that produce HTML graphs of the output.

A RabbitMQ cluster can be limited by a number of factors, from infrastructure-level constraints (e.g. network bandwidth) to RabbitMQ configuration and topology to applications that publish and consume. PerfTest can demonstrate baseline performance of a node or a cluster of nodes.

The most basic way of running PerfTest only specifies a URI to connect to, a number of publishers to use (say, 1) and a number of consumers to use (say, 2). Note that RabbitMQ Java client can achieve high rates for publishing (up to 80 to 90K messages per second per connection), given enough bandwidth and when some safety measures (publisher confirms) are disabled, so overprovisioning publishers is rarely necessary (unless that’s a specific objective of the test).

The following command runs PerfTest with a single publisher without publisher confirms, two consumers (each receiving a copy of every message) that use automatic acknowledgement mode and a single queue named “throughput-test-x1-y2”. Publishers will publish as quickly as possible, without any rate limiting. Results will be prefixed with “test1” for easier identification and comparison:

This modification will use 2 publishers and 4 consumers, typically yielding higher throughput given enough CPU cores on the machine and RabbitMQ nodes:

This modification switches consumers to manual acknowledgements:

This modification changes message size from default (12 bytes) to 4 kB:

PerfTest can use durable queues and persistent messages:

When PerfTest is running, it is important to monitor various publisher and consumer metrics provided by the management UI. For example, it is possible to see how much network bandwidth a publisher has been using recently on the connection page.

Queue page demonstrates message rates, consumer count, acknowledgement mode used by the consumers, consumer utilisation and message location break down (disk, RAM, paged out transient messages, etc). When durable queues and persistent messages are used, node I/O and message store/queue index operation metrics become particularly important to monitor.

Consumers can ack multiple messages at once, for example, 100 in this configuration:

Consumer prefetch (QoS) can be configured as well (in this example to 500):

Publisher confirms can be used with a maximum of N outstanding publishes:

PerfTest can publish only a certain number of messages instead of running until shut down:

Publisher rate can be limited:

Consumer rate can be limited as well to simulate slower consumers or create a backlog:

Note that the consumer rate limit is applied per consumer, so in the configuration above the limit is actually 2 * 2000 = 4000 deliveries/second.

PerfTest can be configured to run for a limited amount of time in seconds with the -z option:

Running PerfTest without consumers and with a limited number of messages can be used to pre-populate a queue, e.g. with 1M messages 1 kB in size each::

Use the -D option to limit the number of consumed messages. Note the -z (time limit), -C (number of published messages), and -D (number of consumed messages) options can be used together but their combination can lead to funny results. -r 1 -x 1 -C 10 -y 1 -D 20 would for example stop the producer once 10 messages have been published, letting the consumer wait forever the remaining 10 messages (as the publisher is stopped).

To consume from a pre-declared and pre-populated queue without starting any publishers, use

PerfTest is useful for establishing baseline cluster throughput with various configurations but does not simulate many other aspects of real world applications. It is also biased towards very simplistic workloads that use a single queue, which provides limited CPU utilisation on RabbitMQ nodes and is not recommended for most cases.

Multiple PerfTest instances running simultaneously can be used to simulate more realistic workloads.

If a queue name is defined (-u "queue-name"), PerfTest will create a queue with this name and all consumers will consume from this queue. The queue will be bound to the direct exchange with its name as the routing key. The routing key will be used by producers to send messages. This will cause messages from all producers to be sent to this single queue and all consumers to receive messages from this single queue.

If the queue name is not defined, PerfTest will create a random UUID routing key with which producers will publish messages. Each consumer will create its own anonymous queue and bind it to the direct exchange with this routing key. This will cause each message from all producers to be replicated to multiple queues (number of queues equals number of consumers), while each consumer will be receiving messages from only one queue.

Note it is possible to customise the queue and to work against several queues as well.

There are 2 reasons for a PerfTest run to stop:

In both cases, PerfTest tries to exit as cleanly as possible, in a reasonable amount of time. Nevertheless, when PerfTest AMQP connections are throttled by the broker, because they’re publishing too fast or because broker alarms have kicked in, it can take time to close them (several seconds or more for one connection).

If closing connections in the gentle way takes too long (5 seconds by default), PerfTest will move on to the most important resources to free and terminates. This can result in client unexpectedly closed TCP connection messages in the broker logs. Note this means the AMQP connection hasn’t been closed with the right sequence of AMQP frames, but the socket has been closed properly. There’s no resource leakage here.

The connection closing timeout can be set up with the --shutdown-timeout argument (or -st). The default timeout can be increased to let more time to close connections, e.g. the command below uses a shutdown timeout of 20 seconds:

The connection closing sequence can also be skipped by setting the timeout to 0 or any negative value:

With the previous command, PerfTest won’t even try to close AMQP connections, it will exit as fast as possible, freeing only the most important resources. This is perfectly acceptable when performing runs on a test environment.

PerfTest can create queues using provided queue arguments:

The previous command will create a queue with a length limit of 10. You can also provide several queue arguments by separating the key/value pairs with commas:

It is possible to use several arguments to create quorum queues, but PerfTest provides a --quorum-queue flag to do that:

--quorum-queue is a shortcut for --flag persistent --queue-args x-queue-type=quorum --auto-delete false. Note a quorum queue cannot have a server-generated name, so the --queue argument must be used to specify the name of the queue(s).

It is possible to customise messages that PerfTest publishes. This allows getting as close as possible to the target traffic or to populate queues with messages that real consumers will process.

By default, PerfTest publishes as fast as possible. The publishing rate per producer can be limited with the --rate option (-r). E.g. to publishing at most 100 messages per second for the whole run:

The --variable-rate (-vr) option can be used several times to specify a publishing rate for a duration, e.g.:

The variable rate option uses the [RATE]:[DURATION] syntax, where RATE is in messages per second and DURATION is in seconds. In the previous example, the publishing rate will be 100 messages per second for 60 seconds, then 1000 messages per second for 10 seconds, then 500 messages per second for 15 seconds, then back to 100 messages per second for 60 seconds, and so on.

The --variable-rate option is useful to simulate steady rates and burst of messages for short periods.

The default size of the messages that PerfTest publishes is 12 bytes (PerfTest stores some data in the message to calculate latency on the consumer side).

It is possible to make messages bigger with the --size (-s) option, e.g. to publish 4 kB messages:

The --variable-size (-vs) option allows to specify different message sizes for periods of time, e.g.:

The variable rate option uses the [SIZE]:[DURATION] syntax, where SIZE is in bytes and DURATION is in seconds. In the previous example, the size of published messages will be 1 kB for 30 seconds, then 10 kB for 20 seconds, then 5 kB for 45 seconds, then back to 1 kB for 30 seconds, and so on.

You can simulate processing time per message with either a fixed or a variable latency value in microseconds.

The --consumer-latency (-L) option sets a fixed consumer latency in microseconds. In the example below a 1 ms latency is set:

The --variable-latency (-vl) option sets a variable consumer latency. In the example below it is set to 1 ms for 60 seconds then 1 second for 30 seconds:

PertTest supports balancing the publishing and the consumption across a sequence of queues, e.g.:

The previous command would create the perf-test-1, perf-test-2, …​, perf-test-10 queues and spreads the producers and consumers across them. This way each queue will have 10 consumers and 10 producers sending messages to it.

Load is balanced in a round-robin fashion:

With the previous command, queues from perf-test-1 to perf-test-5 will have 2 producers, and queues from perf-test-6 to perf-test-10 will have only 1 producer. Each queue will have 3 consumers.

Note the --queue-pattern value is a Java printf-style format string. The queue index is the only argument passed in. The formatting is very close to C’s printf. --queue-pattern 'perf-test-%03d' --queue-pattern-from 1 --queue-pattern-to 500 would for instance create queues from perf-test-001 to perf-test-500.

PerfTest can easily run hundreds of connections on a simple desktop machine. Each producer and consumer use a Java thread and a TCP connection though, so a PerfTest process can quickly run out of file descriptors, depending on the OS settings. A simple solution is to use several PerfTest processes, on the same machine or not. This is especially handy when combined with the queue sequence feature.

The following command line launches a first PerfTest process that creates 500 queues (from perf-test-1 to perf-test-500). Each queue will have 3 consumers and 1 producer sending messages to it:

Then the following command line launches a second PerfTest process that creates 500 queues (from perf-test-501 to perf-test-1000). Each queue will have 3 consumers and 1 producer sending messages to it:

Those 2 processes will simulate 1000 producers and 3000 consumers spread across 1000 queues.

A PerfTest process can exhaust its file descriptors limit and throw java.lang.OutOfMemoryError: unable to create new native thread exceptions. A first way to avoid this is to reduce the number of Java threads PerfTest uses with the --heartbeat-sender-threads option:

By default, each producer and consumer connection uses a dedicated thread to send heartbeats to the broker, so this is 4000 threads for heartbeats in the previous sample. Considering producers and consumers always communicate with the broker by publishing messages or sending acknowledgments, connections are never idle, so using 10 threads for heartbeats for the 4000 connections should be enough. Don’t hesitate to experiment to come up with the appropriate --heartbeat-sender-threads value for your use case.

Another way to avoid java.lang.OutOfMemoryError: unable to create new native thread exceptions is to tune the number of file descriptors allowed per process at the OS level, as some distributions use very low limits. Here the recommendations are the same as for the broker, so you can refer to our networking guide.

A typical connected device workload (a.k.a "IoT workload") involves many producers and consumers (dozens or hundreds of thousands) that exchange messages at a low and mostly constant rate, usually a message every few seconds or minutes. Simulating such workloads requires a different set of settings compared to the workloads that have higher throughput and a small number of clients. With the appropriate set of flags, PerfTest can simulate IoT workloads without requiring too many resources, especially threads.

With an IoT workload, publishers usually don’t publish many messages per second, but rather a message every fixed period of time. This can be achieved by using the --publishing-interval flag instead of the --rate one. For example:

The command above makes the publisher publish a message every 5 seconds. To simulate a group of consumers, use the --queue-pattern flag to simulate many consumers across many queues:

To prevent publishers from publishing at roughly the same time and distribute the rate more evenly, use the --producer-random-start-delay option to add an random delay before the first published message:

With the command above, each publisher will start with a random delay between 1 and 120 seconds.

When using --publishing-interval, PerfTest will use one thread for scheduling publishing for all 50 producers. So 1000 producers should keep 20 threads busy for the publishing scheduling. This ratio can be decreased or increased with the --producer-scheduler-threads options depending on the load and the target environment. Very few threads can be used for very slow publishers:

In the example above, 1000 publishers will publish every 60 seconds with a random start-up delay between 1 second and 30 minutes (1800 seconds). They will be scheduled by only 10 threads (instead of 20 by default). Such delay values are suitable for long running tests.

Another option can be useful when simulating many consumers with a moderate message rate: --consumers-thread-pools. It allows to use a given number of thread pools for all the consumers, instead of one thread pool for each consumer by default. In the previous example, each consumer would use a 1-thread thread pool, which is overkill considering consumers processing is fast and producers publish one message every second. We can set the number of thread pools to use with --consumers-thread-pools and they will be shared by the consumers:

The previous example uses only 10 thread pools for all consumers instead of 1000 by default. These are 1-thread thread pools in this case, so this is 10 threads overall instead of 1000, another huge resource saving to simulate more clients with a single PerfTest instance for large IoT workloads.

By default, PerfTest uses blocking network socket I/O to communicate with the broker. This mode works fine for clients in many cases but the RabbitMQ Java client also supports an asynchronous I/O mode, where resources like threads can be easily tuned. The goal here is to use as few resources as possible to simulate as much load as possible with a single PerfTest instance. In the slow publisher example above, a handful of threads should be enough to handle the I/O. That’s what the --nio-threads flag is for:

This way PerfTest will use 12 threads for I/O over all the connections. With the default blocking I/O mode, each producer (or consumer) uses a thread for the I/O loop, that is 2000 threads to simulate 1000 producers and 1000 consumers. Using NIO in PerfTest can dramatically reduce the resources used to simulate workloads with a large number of connections with appropriate tuning.

Note that in NIO mode the number of threads used can increase temporarily when connections close unexpectedly and connection recovery kicks in. This is due to the NIO mode dispatching connection closing to non-I/O threads to avoid deadlocks. Connection recovery can be disabled with the --disable-connection-recovery flag.

If you run producers and consumers on different machines or even in different processes, and you want PerfTest to calculate latency, you need to use the --use-millis flag. E.g. for sending messages from one host:

And for consuming messages from another host:

Note that as soon as you use --use-millis, latency is calculated in milliseconds instead of microseconds. Note also the different machines should have their clock synchronised, e.g. by NTP. If you don’t run producers and consumers on different machines or if you don’t want PerfTest to calculate latency, you don’t need the --use-millis flag.

Why does one need to care about the --use-millis flag? PerfTest uses by default System.nanoTime() in messages to calculate latency between producers and senders. System.nanoTime() provides nanosecond precision but must be used only in the same Java process. So PerfTest can fall back to System.currentTimeMillis(), which provides only milliseconds precision, but is reliable between different machines as long as their clocks are synchronised.

Consumers are asynchronous by default in PerfTest. This means they are registered with the AMQP basic.consume method and the broker pushes messages to them. This is the optimal way to consume messages. PerfTest also provides the --polling and --polling-interval options to consume messages by polling the broker with the AMQP basic.get method. These options are available to evaluate the performance and the effects of basic.get, but real applications should avoid using basic.get as much as possible because it has several drawbacks compared to asynchronous consumers: it needs a network round trip for each message, it typically keeps a thread busy for polling in the application, and it intrinsically increases latency.

PerfTest can use TLS to connect to a node that is configured to accept TLS connections. To enable TLS, simply specify a URI that uses the amqps schema:

By default PerfTest automatically trusts the server and doesn’t present any client certificate (a warning shows up in the console). In many benchmarking or load testing scenarios this may be sufficient. If peer verification is necessary, it is possible to use the appropriate JVM properties on the command line to override the default SSLContext. For example, to trust a given server:

The previous snippet uses a one-liner to define the JAVA_OPTS environment variable while running PerfTest. Please refer to the TLS guide to learn about how to set up RabbitMQ with TLS. A convenient way to generate a CA and some self-signed certificate/key pairs for development and QA environments is with tls-gen. `tls-gen’s basic profile is a good starting point. How to run PerfTest with a certificate/key pair generated by the aforementioned profile:

Environment variables can sometimes be easier to work with than command line options, for example when using a manifest file to configure PerfTest (with Docker Compose or Cloud Foundry), especially when the number of options used grows.

PerfTest will automatically use environment variables that match the snake case version of the long version of its options (e.g. PerfTest will automatically pick up the value of the CONFIRM_TIMEOUT environment variable for the --confirm-timeout option, but only if the environment variable is defined).

You can list the environment variables that PerfTest will pick up with the following command:

Note that some options can be used several times to define several values, e.g.:

Declaring an environment variable several times just overrides the previous value, so to define several values for an environment variable, just separate the values with a comma:

To avoid collisions with environment variables that already exist, it is possible to specify a prefix for the environment variables that PerfTest will look up. This prefix is defined with the RABBITMQ_PERF_TEST_ENV_PREFIX environment variable, e.g.:

With RABBITMQ_PERF_TEST_ENV_PREFIX="PERF_TEST_" defined, PerfTest will for example look for the PERF_TEST_CONFIRM_TIMEOUT environment variable, not only CONFIRM_TIMEOUT.

PerfTest is also distributed as a native executable built with GraalVM. The native executable has the following advantages: it doesn’t need a JVM to run, it has faster startup time and lower runtime memory overhead compared to a Java VM.

PerfTest native executable has also some limitations:

The PerfTest HTML extension are a set of tools that can help you run automated benchmarks by wrapping around PerfTest. You can provide benchmark specs, and the tool will take care of running the benchmark, collecting results and displaying them in an HTML page. Learn more here.

PerfTest can gather metrics and make them available to various monitoring systems. Metrics include messaging-centric metrics (message latency, number of connections and channels, number of published messages, etc) as well as OS process and JVM metrics (memory, CPU usage, garbage collection, JVM heap, etc).

Here is how to list the available metrics options:

This command displays the available flags to enable the various metrics PerfTest can gather, as well as options to configure the exposure to the monitoring systems PerfTest supports.