Rosbag2 writer benchmarking

The primary package to test performance of the rosbag2.

How it works

Use benchmark_launch.py launchfile to run an entire set of benchmarks.

Launchfile requires two arguments:

• benchmark - provides benchmark description (how many repetitions, cache size, database configuration etc.),
• producers - provides producers description (how many publisher/producer instances, frequency, messages size etc.)

Templates for these configuration files are in config directory of this package.

To run test benchmark (with test.yaml and mixed_110Mbs.yaml):

ros2 launch rosbag2_performance_benchmarking benchmark_launch.py benchmark:=`ros2 pkg prefix rosbag2_performance_benchmarking`/share/rosbag2_performance_benchmarking/config/benchmarks/test.yaml producers:=`ros2 pkg prefix rosbag2_performance_benchmarking`/share/rosbag2_performance_benchmarking/config/producers/mixed_110Mbs.yaml

The summary of benchmark goes into result file described in benchmark config: <bag_root_folder>/<BENCHMARK_NAME>/summary_result_file where BENCHMARK_NAME is a name generated from config names, transport type and timestamp.

For human friendly output, a postprocess report generation tool can be used. Launch it with benchmark result directory as an -i argument (directory with summary_result_file file):

scripts/report_gen.py -i <BENCHMARK_RESULT_DIR>

Binaries

These are used in the launch file:

• benchmark_publishers - runs publishers based on provided parameters. Used when no_transport parameter is set to False;
• writer_benchmark - runs storage-only benchmarking, mimicking subscription queues but using no transport whatsoever. Used when no_transport parameter is set to True.
• results_writer - based on provider parameters, write results (percentage of recorded messages) after recording. One of the parameters is the storage uri, which is used to read the bag metadata file.

Compression

Note that while you can opt to select compression for benchmarking, the generated data is random so it is likely not representative for this specific case. To publish non-random data, you need to modify the ByteProducer.

Number of publisher threads

In the case of the benchmark_publishers binary, a pool of threads is created to run the publishers. By default, the number of threads is equal to the number of publishers. It is possible to change the number of threads using the optional number_of_threads parameter.

Building

To build the package in the rosbag2 build process, make sure to turn BUILD_ROSBAG2_BENCHMARKS flag on (e.g. colcon build --cmake-args -DBUILD_ROSBAG2_BENCHMARKS=1)

If you already built rosbag2, you can use packages-select option to build benchmarks. Example: colcon build --packages-select rosbag2_performance_benchmarking --cmake-args -DBUILD_ROSBAG2_BENCHMARKS=1.

General knowledge: I/O benchmarking

Background: benchmarking disk writes on your system

It might be useful to first understand what limitation your disk poses to the throughput of data recording. Performance of bag write can’t be higher over extended period of time (you can only use as much memory).

Using dd command

dd if=/dev/zero of=/tmp/output conv=fdatasync bs=384k count=1k; rm -f /tmp/output

This method is not great for benchmarking the disk but an easy way to start since it requires no dependencies. This will write zeros to the /tmp/output file with block size 384k, 1000 blocks, ends when write finishes. Make sure to benchmark the disk which your bags will be written to (check your mount points and change “/tmp/output” to another path if needed). Note: this depends on parameters used and whatever else is running on your system but can give you a ballpark figure when ran several times.

Using fio

For more sophisticated & accurate benchmarks, see the fio command. An example for big data blocks is: fio --name TEST --eta-newline=5s --filename=fio-tempfile.dat --rw=write --size=500m --io_size=10g --blocksize=1024k --ioengine=libaio --fsync=10000 --iodepth=32 --direct=1 --numjobs=1 --runtime=60 --group_reporting.

Profiling bags I/O with tools

Tools that can help in I/O profiling: sudo apt-get install iotop ioping sysstat

• iotop works similar as top command, but shows disk reads, writes, swaps and I/O %. Can be used at higher frequency in batch mode with specified process to deliver data that can be plotted.
  • Example use: sudo iotop -h -d 0.1 -t -b -o -p <PID> after running the bag.
• ioping can be used to check latency of requests to device
• strace can help determine syscalls associated with the bottleneck.
  • Example use: strace -c ros2 bag record /image --max-cache-size 10 -o ./tmp. You will see a report after finishing recording with Ctrl-C.

Changelog for package rosbag2_performance_benchmarking

0.26.5 (2024-09-06)

0.26.4 (2024-06-27)

0.26.3 (2024-05-15)

0.26.2 (2024-04-24)

0.26.1 (2024-04-17)

0.26.0 (2024-04-16)

• Use middleware send and receive timestamps from message_info during recording (#1531)
• Update to use yaml-cpp version 0.8.0. (#1605)
• Contributors: Chris Lalancette, jmachowinski

0.25.0 (2024-03-27)

• Add option to set compression threads priority (#1457)
• Add per group statistics for rosbag2_performance_benchmarking report (#1306)
• Contributors: Michael Orlov, jmachowinski

0.24.0 (2023-07-11)

• Set CPU affinity for producers and recorder from benchmark parameters (#1305)
• Add CPU usage to rosbag2_performance_benchmarking results report (#1304)
• Add config option to use storage_id parameter in benchmark_launch.py (#1303)
• Contributors: Michael Orlov

0.23.0 (2023-04-28)

0.22.0 (2023-04-18)

• Add tests for rosbag2_performance_benchmarking pkg (#1268)
• Contributors: Michael Orlov

0.21.0 (2023-04-12)

• Fix expectations for rosbag2 return code in rosbag2_performance_benchmarking (#1267)
• Update rosbag2 to C++17. (#1238)
• Use thread pool to run benchmark publishers in rosbag2_performance_benchmarking (#1250)
• Use target_link_libraries instead of ament_target_dependencies (#1202)
• Contributors: Chris Lalancette, Daisuke Nishimatsu, Michael Orlov, Shane Loretz, carlossvg

0.20.0 (2023-02-14)

• Skip ament_package() call when not building rosbag2_performance_benchmarking (#1242)
• Contributors: Shane Loretz

0.19.0 (2023-01-13)

• Add option to specify a message type (#1153)
• Add Michael Orlov as maintainer in rosbag2 packages (#1215)
• Replace language for "db3"/"db"/"database" (#1194)
• Remove explicit sqlite3 from code (#1166)
• Contributors: Emerson Knapp, Michael Orlov, carlossvg

0.18.0 (2022-11-15)

0.17.0 (2022-07-30)

0.16.0 (2022-05-11)

0.15.1 (2022-04-06)

0.15.0 (2022-04-05)

0.14.1 (2022-03-29)

• Bump version number to avoid conflict
• Contributors: Chris Lalancette

0.14.0 (2022-03-29)

• Install headers to include/${PROJECT_NAME} (#958)
• Contributors: Shane Loretz

0.13.0 (2022-01-13)

0.12.0 (2021-12-17)

• Enable YAML encoding/decoding for RecordOptions and StorageOptions (#916)
  • Enable YAML encoding/decoding for RecordOptions and StorageOptions
• Contributors: Emerson Knapp

0.11.0 (2021-11-08)

• Update package maintainers (#899)
• Contributors: Michel Hidalgo

0.10.1 (2021-10-22)

0.10.0 (2021-10-19)

• Updated node declare_parameter to new syntax (#882)
• Updated benchmark package to use writer close() instead of old reset() (#881)
• Contributors: Adam Dąbrowski

0.9.0 (2021-05-17)

0.8.0 (2021-04-19)

0.7.0 (2021-03-18)

• fixed a memory leak in no-transport benchmark (#674)
• report of performance improvements in rosbag2 (roughly since Foxy) (#651)
• Performance benchmarking improvements (#634)
• Contributors: Adam Dąbrowski, Piotr Jaroszek

0.6.0 (2021-02-01)

• Performance benchmarking refactor (#594)
• Contributors: Adam Dąbrowski

0.5.0 (2020-12-02)

• Sqlite storage double buffering (#546)
• Contributors: Adam Dąbrowski

0.4.0 (2020-11-19)

• read yaml config file (#497)
• add storage_config_uri (#493)
• Update the package.xml files with the latest Open Robotics maintainers (#535)
• performance testing packages (#442)
• Contributors: Adam Dąbrowski, Karsten Knese, Michael Jeronimo

0.3.2 (2020-06-03)

0.3.1 (2020-06-01)

0.3.0 (2020-05-26)

0.2.8 (2020-05-18)

0.2.7 (2020-05-12)

0.2.6 (2020-05-07)

0.2.5 (2020-04-30)

0.2.4 (2019-11-18 17:51)

0.2.3 (2019-11-18 13:55)

0.2.2 (2019-11-13)

0.2.1 (2019-10-23)

0.2.0 (2019-09-26)

0.1.2 (2019-05-20)

0.1.1 (2019-05-09)

0.1.0 (2019-05-08)

0.0.5 (2018-12-27)

0.0.4 (2018-12-19)

0.0.3 (2018-12-14)

0.0.2 (2018-12-12)

0.0.1 (2018-12-11)