rosbag2_performance_benchmarking

Code to benchmark rosbag2

README

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: <db_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.

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.