Skip to content

Ring buffer that allows for high-throughput data transfer between multiproccessing Python processes.

License

Notifications You must be signed in to change notification settings

bslatkin/ringbuffer

Repository files navigation

A multiple writer and reader ring buffer that allows for high-throughput data transfer between multiproccessing Python processes. The goal is to make it easy to construct bandwidth-heavy pipelines (e.g., for video stream processing) that can utilize multiple cores with Python tools like OpenCV, Scikit Learn, and TensorFlow.

The RingBuffer data structure's performance is primarily bound by the behavior of the Lock class, which is a Kernel semaphore under the covers. The lock is used to coordinate a readers/writer lock, meaning that lock contention dominates as the number of writes per second increases. Memory performance isn't an issue because all data is transferred through mmap'ed buffers.

For examples of how it all fits together, look at example_numpy.py and example_ctypes.py.

On an old MacBook, the ring buffer can easily do 2 gigabytes per second of transfer when using large slot sizes (~100MB), a relatively low number of writes per second (~24), and a single reader/writer pair. As you increase the number of writes per second the performance degrades proportionately due to the lock. However, increasing the number of readers has a minimal effect (when the slot sizes are ~10MB) because multiple readers can hold the read lock at the same time.

Build Status

Works with Python 3.5 and later. Background on why it doesn't work with Python 3.3 and 3.4 is here.


Included is a tool called perf_test_ringbuffer.py that tests the performance characteristics of the ring buffer. This can be used to exercise the various ways in which the RingBuffer scales, including number of readers, number of writes per second, slow readers, etc.

Example that shows good behavior:

./perf_test_ringbuffer.py \
    --debug \
    --slot-bytes=1000000 \
    --slots=50 \
    --duration-seconds=10 \
    --writes-per-second=24 \
    --readers=5

Example that shows that too many readers will slow the system down due to lock contention:

./perf_test_ringbuffer.py \
    --debug \
    --slot-bytes=1000000 \
    --slots=10 \
    --duration-seconds=10 \
    --writes-per-second=24 \
    --readers=100

Example that shows how the writer will fall behind its target rate when the locking overhead becomes too large:

./perf_test_ringbuffer.py \
    --debug \
    --slot-bytes=1000000 \
    --slots=10 \
    --duration-seconds=10 \
    --writes-per-second=2000 \
    --readers=1

Example that shows how the writer will fall behind its target rate when the requested data transfer rate is too high for the memory performance of the machine:

./perf_test_ringbuffer.py \
    --debug \
    --slot-bytes=100000000 \
    --slots=3 \
    --duration-seconds=10 \
    --writes-per-second=24 \
    --readers=1 \
    --no-verify_writes

Example that shows what happens when the readers can't keep up with the writer:

./perf_test_ringbuffer.py \
    --debug \
    --slot-bytes=1000000 \
    --slots=10 \
    --duration-seconds=3 \
    --writes-per-second=24 \
    --readers=4 \
    --reader-burn-cpu-milliseconds=100

About

Ring buffer that allows for high-throughput data transfer between multiproccessing Python processes.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published