Add receiver-side IPC benchmark for Image::poll() / Subscription::poll()

[benkier0]

I'm looking to make a first contribution and identified what seems like a gap in the current C++ benchmark coverage in aeron-io/benchmarks.

All four existing C++ benchmarks (AeronIpcBenchmark, AeronExclusiveIpcBenchmark, and their Nanomark variants) measure publisher throughput only. There's no performance baseline for the receiver hot-path, Image::poll() and Subscription::poll(), making it impossible to profile end to end publish-to-consume latency over IPC from the C++ benchmarks alone

My plan:

• mirror the structure of AeronIpcNanomark.cpp
• IPC transport, background publisher thread, measure fragment delivery throughput/latency from the polling side
• Register via the existing benchmark() CMake macro in aeron-io/benchmarks
• HDR histogram output consistent with the existing nanomark pattern

Does this approach fit what you'd want? Should I also include Image::controlledPoll() in the same PR, or keep the first PR focused on the basic poll path?

[pveentjer]

If you are planning to measure latency, do you want to measure latency at the highest throughput possible? Or will there be some fixed rate of requests and then measure latency based on the intended start time of a request instead of the actual start time of the request (to deal with coordinated omission).

If you do not have a fixed rate of request, the primary sensible latency metric would be the average latency. The percentiles would be pretty useless due to coordinated omission.

[benkier0]

Fixed rate with intended-start-time measurement, so latency is wall-clock(receive) minus intended_send_time, not actual_send_time. This makes p99/p999 meaningful. Id expose rate as a configurable parameter so users can benchmark at their operating point rather than just saturation. Does that align with what you'd want, or is there a preferred pattern in the existing benchmarks I should mirror?

[pveentjer]

Good to see that we are on the same page.

We use the same approach in the Aeron benchmarks that tests the whole system. You can check out the LoadTestRig.send method for some ideas.

[benkier0]

I created a pr. aeron-io/benchmarks#110