performance_test

[TOC]

The performance_test tool tests latency and other performance metrics of various middleware implementations that support a pub/sub pattern. It is used to simulate non-functional performance of your application.

The performance_test tool allows you to quickly set up a pub/sub configuration, e.g. number of publisher/subscribers, message size, QOS settings, middleware. The following metrics are automatically recorded when the application is running:

• latency: corresponds to the time a message takes to travel from a publisher to subscriber. The latency is measured by timestamping the sample when it's published and subtracting the timestamp (from the sample) from the measured time when the sample arrives at the subscriber (only logged when a subscriber is created)
• CPU usage: percentage of the total system wide CPU usage (logged separately for each instance of perf_test)
• resident memory: heap allocations, shared memory segments, stack (used for system's internal work) (logged separately for each instance of perf_test)
• sample statistics: number of samples received, sent, and lost per experiment run.

This master branch is compatible with the following ROS 2 versions - rolling - humble - galactic - foxy - eloquent - dashing - Apex.OS

How to use this document

1. Start here for a quick example of building and running the performance_test tool with the Cyclone DDS plugin.
2. If needed, find more detailed information about building and running
3. Or, if the quick example is good enough, skip ahead to the list of supported middleware plugins to learn how to test a specific middleware implementation.
4. Check out the tools for visualizing the results
5. If desired, read about the design and architecture of the tool.

Example

This example shows how to test the non-functional performance of the following configuration:

1. Install ROS 2

2. Install Cyclone DDS to /opt/cyclonedds

3. Build performance_test with the CMake build flag for Cyclone DDS:

    source /opt/ros/rolling/setup.bash
    cd ~/perf_test_ws
    colcon build --cmake-args -DPERFORMANCE_TEST_PLUGIN=CYCLONEDDS
    source ./install/setup.bash

1. Run with the communication plugin option for Cyclone DDS:

mkdir experiment
./install/performance_test/lib/performance_test/perf_test --communication CycloneDDS
                                                          --msg Array1k
                                                          --rate 100
                                                          --topic test_topic
                                                          --max-runtime 30
                                                          --logfile experiment/log.csv

At the end of the experiment, a CSV log file will be generated in the experiment folder with a name that starts with log.

Building the performance_test tool

For a simple example, see Dockerfile.rclcpp.

The performance_test tool is structured as a ROS 2 package, so colcon is used to build it. Therefore, you must source a ROS 2 installation:

source /opt/ros/rolling/setup.bash

Select a middleware plugin from this list. Then build the performance_test tool with the selected middleware:

mkdir -p ~/perf_test_ws/src
cd ~/perf_test_ws/src
git clone https://gitlab.com/ApexAI/performance_test.git
cd ..
# At this stage, you need to choose which middleware you want to use
# The list of available flags is described in the middleware plugins section
# Square brackets denote optional arguments, like in the Python documentation.
colcon build --cmake-args -DCMAKE_BUILD_TYPE=Release -DPERFORMANCE_TEST_PLUGIN=<plugin>
source install/setup.bash

Running an experiment

The performance_test experiments are run through the perf_test executable. To find the available settings, run with --help (note the required and default arguments):

~/perf_test_ws$ ./install/performance_test/lib/performance_test/perf_test --help

• The -c argument should match the selected middleware plugin from the build phase.
• The --msg argument should be one of the supported message types, which are shown in the --help output.

Single machine or distributed system?

Based on the configuration you want to test, the usage of the performance_test tool differs. The different possibilities are explained below.

For running tests on a single machine, you can choose between the following options:

1. Intraprocess means that the publisher and subscriber threads are in the same process.

    perf_test <options> --num-sub-threads 1 --num-pub-threads 1

1. Interprocess means that the publisher and subscriber are in different processes. To test interprocess communication, two instances of the performance_test must be run, e.g.

    # Start the subscriber first
    perf_test <options> --num-sub-threads 1 --num-pub-threads 0 &
    sleep 1  # give the subscriber time to finish initializing
    perf_test <options> --num-sub-threads 0 --num-pub-threads 1

On a distributed system, testing latency is difficult, because the clocks are probably not perfectly synchronized between the two devices. To work around this, the performance_test tool supports relay mode, which allows for a round-trip style of communication:

# On the main machine
perf_test <options> --roundtrip-mode Main

# On the relay machine:
perf_test <options> --roundtrip-mode Relay

In relay mode, the Main machine sends messages to the Relay machine, which immediately sends the messages back. The Main machine receives the relayed message, and reports the round-trip latency. Therefore, the reported latency will be roughly double the latency compared to the latency reported in non-relay mode.

Single machine, single thread

An intra-thread configuration is experimentally supported, in which a publisher and subscriber both operate in the same thread. The publisher writes a messages, and the subscriber immediately takes it.

perf_test <options> -e INTRA_THREAD

Notes:

1. This is only available when zero copy transfer is enabled
2. This requires exactly one publisher and one subscriber
3. This is not compatible with roundtrip mode

Middleware plugins

The performance test tool can measure the performance of a variety of communication solutions from different vendors. In this case there is no rclcpp or rmw layer overhead over the publisher and subscriber routines.

The performance_test tool implements an executor that runs the publisher(s) and/or the subscriber(s) in their own thread.

The following plugins are currently implemented:

Eclipse Cyclone DDS

• Eclipse Cyclone DDS 0.9.0b1
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=CYCLONEDDS
• Communication plugin: -c CycloneDDS
• Docker file: Dockerfile.CycloneDDS
• Available transports:
  • Cyclone DDS zero copy requires the runtime switch to be enabled.
  • When the runtime switch is enabled, RouDi must be running.
  • If the runtime switch is enabled, but --zero-copy is not added, then the plugin will not use the loaned sample API, but iceoryx will still transport the samples.

  - See Dockerfile.mashup

  Pub/sub in same process	Pub/sub in different processes on same machine	Pub/sub in different machines
  INTRA (default), LoanedSamples (--zero-copy)	UDP (default), LoanedSamples (--zero-copy), SHMEM (enable runtime switch)	UDP

Eclipse Cyclone DDS C++ binding

• Eclipse Cyclone DDS C++ bindings 0.9.0b1
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=CYCLONEDDS_CXX
• Communication plugin: -c CycloneDDS-CXX
• Docker file: Dockerfile.CycloneDDS-CXX
• Available transports:
  • Cyclone DDS zero copy requires the runtime switch to be enabled.
  • When the runtime switch is enabled, RouDi must be running.
  • If the runtime switch is enabled, but --zero-copy is not added, then the plugin will not use the loaned sample API, but iceoryx will still transport the samples.

  - See Dockerfile.mashup

  Pub/sub in same process	Pub/sub in different processes on same machine	Pub/sub in different machines
  INTRA (default), LoanedSamples (--zero-copy), SHMEM (enable runtime switch)	UDP (default), LoanedSamples (--zero-copy), SHMEM (enable runtime switch)	UDP

Eclipse iceoryx

• iceoryx (latest master as of Feb 13)
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=ICEORYX
• Communication plugin: -c iceoryx
• Docker file: Dockerfile.iceoryx
• The iceoryx plugin is not a DDS implementation.
  • The DDS-specific options (such as domain ID, durability, and reliability) do not apply.
• To run with the iceoryx plugin, RouDi must be running.
• Available transports: | Pub/sub in same process | Pub/sub in different processes on same machine | Pub/sub in different machines | |-----------|---------------------|-----------------------------------| | LoanedSamples | LoanedSamples | Not supported by performance_test |

eProsima Fast DDS

• FastDDS 2.6.2
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=FASTDDS
• Communication plugin: -c FastRTPS
• Docker file: Dockerfile.FastDDS
• Available transports: | Pub/sub in same process | Pub/sub in different processes on same machine | Pub/sub in different machines | |-------|---------------------|--------------------| | INTRA (default), LoanedSamples (--zero-copy) | SHMEM (default), LoanedSamples (--zero-copy) | UDP |

OCI OpenDDS

• OpenDDS 3.13.2
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=OPENDDS
• Communication plugin: -c OpenDDS
• Docker file: Dockerfile.OpenDDS
• Available transports: | Pub/sub in same process | Pub/sub in different processes on same machine | Pub/sub in different machines | |-------|---------------------|--------------------| | TCP | TCP | TCP |

RTI Connext DDS

• RTI Connext DDS 5.3.1+
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=CONNEXTDDS
• Communication plugin: -c ConnextDDS
• Docker file: Not available
• A license is required
• You need to source an RTI Connext DDS environment.
  • If RTI Connext DDS was installed with ROS 2 (Linux only):
  • source /opt/rti.com/rti_connext_dds-5.3.1/setenv_ros2rti.bash
  • If RTI Connext DDS is installed separately, you can source the following script to set the environment:
  • source <connextdds_install_path>/resource/scripts/rtisetenv_<arch>.bash
• Available transports: | Pub/sub in same process | Pub/sub in different processes on same machine | Pub/sub in different machines | |-------|---------------------|--------------------| | INTRA | SHMEM | UDP |

RTI Connext DDS Micro

• Connext DDS Micro 3.0.3
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=CONNEXTDDSMICRO
• Communication plugin: -c ConnextDDSMicro
• Docker file: Not available
• A license is required
• Available transports: | Pub/sub in same process | Pub/sub in different processes on same machine | Pub/sub in different machines | |-------|---------------------|--------------------| | INTRA | SHMEM | UDP |

Framework plugins

The performance_test tool can also measure the end-to-end latency of a framework. In this case, the executor of the framework is used to run the publisher(s) and/or the subscriber(s). The potential overhead of the rclcpp or rmw layer is measured.

ROS 2

The performance test tool can also measure the performance of a variety of RMW implementations, through the ROS 2 rclcpp::publisher and rclcpp::subscriber API.

• ROS 2 rclcpp::publisher and rclcpp::subscriber
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=ROS2 (default)
• Communication plugin:
  • Callback with Single Threaded Executor: -c rclcpp-single-threaded-executor
  • Callback with Static Single Threaded Executor: -c rclcpp-static-single-threaded-executor
  • rclcpp::WaitSet: -c rclcpp-waitset
• Docker file: Dockerfile.rclcpp
• Available underlying RMW implementations:
  • ROS 2 Rolling is pre-configured to use rmw_fastrtps_cpp
  • Follow these instructions to use a different RMW implementation
• Available transports: depends on underlying RMW implementation
  • LoanedSamples are available (--zero-copy) for ROS_DISTRO = foxy and above

Apex.OS

• Apex.OS
• CMake build flag: -DPERFORMANCE_TEST_PLUGIN=APEX_OS
  • It is also required to source /opt/ApexOS/setup.bash instead of a ROS 2 distribution
• Communication plugin: -c ApexOSPollingSubscription
• Docker file: Not available
• Available underlying RMW implementations: rmw_apex_middleware
• Available transports: | Pub/sub in same process | Pub/sub in different processes on same machine | Pub/sub in different machines | |-------|---------------------|--------------------| | UDP (default), LoanedSamples (--zero_copy), SHMEM (configurable with Apex.OS) | UDP (default), LoanedSamples (--zero_copy), SHMEM (configurable with Apex.OS) | UDP |

Analyze the results

After an experiment is run with the -l flag, a log file is recorded. Both CSV and JSON formats are supported. It is possible to add custom data to the log file by setting theAPEX_PERFORMANCE_TEST environment variable before running an experiment, e.g.

# JSON format
export APEX_PERFORMANCE_TEST="
{
\"My Version\": \"1.0.4\",
\"My Image Version\": \"5.2\",
\"My OS Version\": \"Ubuntu 16.04\"
}
"

Plot results

The performance_test tool provides several tools to plot the generated results:

1. Results rendered on a PDF file: handy to share results
2. Results rendered in a Jupyter notebook: used to compare multiple experiments

Installation

The plot tool requires python3 and texlive. On an Ubuntu system you will need to install the following packages:

sudo apt-get install python3 python3-pip texlive texlive-pictures texlive-luatex texlive-latex-extra

Start a Python virtual environment:

python3 -m venv venv
source venv/bin/activate

Install the required Python packages in that virtual environment:

cd performance_test/helper_scripts/apex_performance_plotter
pip3 install wheel
pip3 install .

Usage

To generate a PDF from the log file, invoke the perfplot binary installed in the previous step:

perfplot <filename1> <filename2> ...

Be sure to also check perfplot -h for additional options.

:point_up: Common Pitfalls

All of the latency metrics are collected and calculated by the subscriber process. For interprocess communication, it is recommended to provide different prefixes for the log files:

perf_test -c rclcpp-single-threaded-executor --msg Array1k -p 0 -s 1 -l log_sub.csv
perf_test -c rclcpp-single-threaded-executor --msg Array1k -p 1 -s 0 -l log_pub.csv

Then, to plot the latency metrics, invoke perfplot on the subscriber's log file. If perfplot is invoked on the publisher's log file, then the CPU and memory metrics will be plotted, but the latency plot will be empty.

To analyze the results in a Jupyter notebook run the following commands:

pipenv shell
jupyter notebook plot_logs.ipynb

# When you are done, deactivate the venv
deactivate

Architecture

Apex.AI's Performance Testing in ROS 2 white paper (available here) describes how to design a fair and unbiased performance test, and is the basis for this project.

Each middleware has a different API. Thanks to the Plugin abstraction, the core logic of setting up and running an experiment is completely decoupled from the implementation details of sending and receiving individual messages.

Exactly one Plugin implementation is selected at build time. The design is similar to the Abstract Factory pattern. performance_test declares, but does not define, a static factory method in the PluginFactory class. Each middleware provides a definition for this factory method to create a concrete Plugin implementation, and perf_test calls this factory method directly.

An example plugin is available here.

Performance optimizations

• On linux-based platforms, perf_test writes 0 to /dev/cpu_dma_latency and holds open the file handle, which will prevent the CPU from entering any idle states for the duration of the experiment. This should result in lower message latency and lower variance in that latency.

Future extensions and limitations

• Communication frameworks like DDS have a huge amount of settings. This tool only allows the most common QOS settings to be configured. The other QOS settings are hardcoded in the application.
• Only one publisher per topic is allowed, because the data verification logic does not support matching data to the different publishers.
• Some communication plugins can get stuck in their internal loops if too much data is received. Figuring out ways around such issues is one of the goals of this tool.
• FastRTPS wait-set does not support timeouts which can lead to the receiving not aborting. In that case the performance test must be manually killed.
• Using Connext DDS Micro INTRA transport with reliable QoS and history kind set to keep_all is not supported with Connext Micro. Set keep-last as QoS history kind always when using reliable.

Possible additional communication which could be implemented are:

• Raw UDP communication

Building with limited resources

When building this tool, the compiler must perform a lot of template expansion. This can be overwhelming for a system with a low-power CPU or limited RAM. There are some additional CMake options which can reduce the system load during compilation:

1. This tool includes many different message types, each with many different sizes. Reduce the number of messages, and thus the compilation load, by disabling one or more message types. For example, to build without PointCloud messages, add -DENABLE_MSGS_POINDCLOUD=OFF to the --cmake-args. The message types, and their options for enabling/disabling, can be found here.

Changelog for package performance_test

X.Y.Z (YYYY/MM/DD)

2.1.0 (2024/04/17)

Added

• Add new function prepare() to the Publisher and Subscriber API, intended to allow participant discovery without blocking the main thread ### Changed
• Change the default --history arg from KEEP_ALL to KEEP_LAST
• Change the default --history-depth arg from 1000 to 16
• If --expected-num-pubs is unspecified, set it to the same value as -p
• If --expected-num-subs is unspecified, set it to the same value as -s ### Fixed
• Removed an unused variable to fix a Clang build
• Remove unused variable names in the Plugin abstract class
• Fix a potential lockup in PublisherTask on QNX

2.0.0 (2024/03/19)

Added

• Add experimental bazel support
  • bazel build //performance_test --//:plugin_implementation=//path/to/a/plugin
• Add a rudimentary socket-based plugin for testing the bazel support
  • bazel run //performance_test --//:plugin_implementation=//performance_test/plugins/demo:demo_plugin -- --help ### Changed
• Instead of enabling/disabling each plugin, you select exactly one with a CMake string option, for example:
  • colcon build --cmake-args -DPERFORMANCE_TEST_PLUGIN=ROS2
• Renamed the --communication CLI arg to --communicator. The short -c is unchanged. ### Removed
• Removed the deprecated CLI flags for QOS settings:
  • Instead of --reliable, use --reliability RELIABLE
  • Instead of --transient, use --durability TRANSIENT_LOCAL
  • Instead of --keep-last, use --history KEEP_LAST
• Removed the obsolete BoundedSequenceFlat messages
• Removed the superfluous --msg-list CLI flag. The --help message already lists the available messages. ### Fixed
• Update the Apex.OS Runner to use executor_runner::deferred instead of executor_runner::deferred_tag()
• Ensure that the first few published samples are sent at the expected rate

1.5.2 (YYYY/MM/DD)

Added

• --prevent-cpu-idle is available on QNX ### Changed
• JSON log files will contain all values in the APEX_PERFORMANCE_TEST dictionary, instead of the five specific values used previously
• Switch to build as C++17 by default ### Fixed
• Zero copy transfer is again enabled for the rclcpp publisher

1.5.0 (2023/06/14)

Added

• New CLI switch --prevent-cpu-idle (linux only). When specified, perf_test will use /dev/cpu_dma_latency to request that the CPU not enter any sleep states, to potentially give more consistent results
• Some smaller Array messages, down to 32 bits
• Added support to the FastDDS plugin for bounded and unbounded sequences ### Changed
• Update the README to better explain how to use this tool with Apex.OS
• In the Runner, allocate the AnalysisResults on the stack instead of using shared_ptr
• Subscriber methods accept a callback parameter, instead of returning a vector of results, to reduce heap usage
• Refactored the interaction between SubscriberStats and AnalysisResult to remove the need for a std::vector of latency samples, to reduce heap usage
• Adjusted the Array message sizes to make the name match the contents
• Updated apex_os_communicator to use the new zero-copy API

1.4.2 (2023/03/15)

Added

• Added perfplot support for JSON log files ### Changed
• Migrate the Apex.OS target to use rosidl_get_typesupport_target
• Preallocate the JSON logger's string buffer to prevent reallocations after the experiment begins

1.4.1 (2023/02/23)

Changed

• Updated the iceoryx plugin to the latest master as of Feb 13

1.4.0 (2023/02/20)

Added

• New message type BoundedSequenceFlat
  • This is a BoundedSequence with the @flat annotation
  • Sizes range from 1kB to 8MB, like Array and BoundedSequence ### Changed
• Messages of different types can be optionally included via CMake args:
  • -DENABLE_MSGS_ARRAY (default ON)
  • -DENABLE_MSGS_STRUCT (default ON)
  • -DENABLE_MSGS_POINT_CLOUD (default ON)
  • -DENABLE_MSGS_BOUNDED_SEQUENCE (default OFF)
  • -DENABLE_MSGS_BOUNDED_SEQUENCE_FLAT (default OFF)
  • -DENABLE_MSGS_UNBOUNDED_SEQUENCE (default OFF)
  • -DENABLE_MSGS_ALL (default OFF)
    • when ON, overrides the other defaults to ON
    • you can still optionally exclude some messages by explicitly setting them to OFF ### Removed
• Removed a few messages:
  • Range
  • RadarTrack
  • RadarDetection
  • NavSatFix ### Fixed
• In all cases, including loaned messages, capture the timestamp as the last step of initializing the message

1.3.7 (2023/01/04)

1.3.6 (2023/01/03)

Fixed

• Set the correct IDL_GEN_ROOT for rclcpp plugins

1.3.5 (2022/12/05)

Fixed

• Exit cleanly when a publisher process terminates before a subscriber process

1.3.4 (2022/11/28)

Changed

• Updated Apex.OS plugins to use the unified LoanedSample::data()

1.3.3 (2022/11/28)

Fixed

• Implement the missing take() method in ApexOSPollingSubscriptionSubscriber

1.3.2 (2022/11/21)

Fixed

• Capture the this pointer in the lambda in the iceoryx publisher

1.3.1 (2022/11/21)

Added

• New Apex.OS plugin, compatible with the ThreadedRunners
  • The INTER_THREAD and INTRA_THREAD execution strategies, combined with -c ApexOSPollingSubscription, will use the ThreadedRunner instances
  • The new APEX_SINGLE_EXECUTOR execution strategy will add all publishers and subscribers to a single Apex.OS Executor
  • The new APEX_EXECUTOR_PER_COMMUNICATOR execution strategy will add each publisher and each subscriber to its own Apex.OS Executor instance
  • The new APEX_CHAIN execution strategy will add a publisher and subscriber as a chain of nodes to an Apex.OS Executor ### Changed
• Refactored FastRTPS communicator plugin:
  • Uses DDS compliant API
  • Code generator updated
  • Implementation for publish_loaned()
  • Dockerfile improvements ### Removed
• CLI arg --disable-async. Synchronous / asynchronous publishing should be configured externally depending on the communication mean used.

1.3.0 (2022/08/25)

Added

• New execution strategy option:
  • The default -e INTER_THREAD runs each publisher and subscriber in its own separate thread, which matches the previous behavior
  • A new -e INTRA_THREAD, which runs a single publisher and subscriber in the same thread. The publisher writes, and the subscriber immediately takes it
  • For Apex.OS specifically, some optimized execution strategies which use the proprietary Apex.OS executor ### Changed
• Significantly refactored the communicator plugins:
  • Each plugin is split into an implementation of a Publisher and a Subscriber, instead of a single Communicator
  • The plugin is no longer responsible for managing the metrics, such as sample count, lost samples, and latency
  • The plugin does not require any special logic to support roundtrip mode
  • It is safe for the plugins to initialize their data writers and readers at construction time, instead of delaying the initialization to the first call of publish() or update_subscription()
  • Split publish() into publish_copy() and publish_loaned()
• Significantly refactored the runner framework:
  • The runner framework is responsible for the experiment metrics
  • It manages the roundtrip mode logic
  • It is extensible for different execution strategies or thread configurations
• The iceoryx plugin now uses the untyped API, for improved performance

1.2.1 (2022/06/30)

Fixed

• Capture the timestamp as soon as a message is received, instead of just before storing the metrics, to reduce the reported latency to a more correct value

1.2.0 (2022/06/28)

Changed

• The CLI arguments for specifying the output type have changed:
  • For console output, updated every second, add --print-to-console
  • For file output, use --logfile my_file.csv or --logfile my_file.json
    • The type will be deduced from the file name
  • If neither of these options is specified, then a warning will print, and the experiment will still run
• The linter configurations are now configured locally. This means that the output of colcon test should be the same no matter the installed ROS distribution.
• The --zero-copy arg is now valid even if the publisher and subscriber(s) are in the same process ### Removed
• The publisher and subscriber loop reserve metrics are no longer recorded or reported ### Fixed
• CPU usage will no longer be stuck at 0

Removed

• The pub/sub loop reserve time metrics

1.1.2 (2022/06/08)

Changed

• Use steady_clock for all platforms, including QNX QOS

1.1.1 (2022/06/07)

Changed

• Significant refactor to simplify the analysis pipeline ### Fixed
• Add some missing definitions when Apex.OS is enabled, but the rclcpp plugins are disabled

1.1.0 (2022/06/02)

Added

• New Apex.OS Polling Subscription plugin
• Compatibility with ROS2 Humble

1.0.0 (2022/05/12)

Added

• More expressive perf_test CLI args for QOS settings
• A plugin for Cyclone DDS with C++ bindings v0.9.0b1 ### Changed
• CLI args for QOS settings:
  • --reliability <RELIABLE|BEST_EFFORT>
  • --durability <TRANSIENT_LOCAL|VOLATILE>
  • --history <KEEP_LAST|KEEP_ALL>
• master branch is compatible with many ROS2 distributions:
  • dashing
  • eloquent
  • foxy
  • galactic
  • rolling ### Deprecated
• CLI flags for QOS settings:
  • --reliable
  • --transient
  • --keep-last ### Removed
• The branches for specific ROS2 distributions have been deleted ### Fixed
• CI jobs and Dockerfiles are decoupled from the middleware bundled with the ROS2 distribution