Qh3 performance benchmark example

[gmarzot]

This is a feature request but I imagine would also be useful as part of the project ci/cd process. To see if any changes help or hurt end to end performance.

[gmarzot]

I have been trying to push/pull higher bandwidth through qh3 .. I suspect I can speed things up with a custom reader but right now i am just processing quic_event_received and end up copying event data and buffered data more than would be ideal.. however this py-spy output seems to show the qh3 layer is consuming the most cpu ... this is single threaded on a Intel(R) Core(TM) i7-9700K CPU @ 3.60GHz and the consuming side is almost at 100% cpu. I am running with 4 quic streams and only getting around 256Mbps over loopback. Much less than possible/expected.

Thank you for any thoughts or requests for additional data

[gmarzot]

moqt/quic publisher profile

moqt/quic subscriber profile

this is py-spy speedscope capture for my publisher and subscriber

can be viewed here: https://speedscope.app

[Ousret]

I'll see what I can do to improve the bottlenecks.

regards,

[gmarzot]

another observation is that when i run my publisher (primarily calling create_wt_stream() send_stream_data() transmit() ), the cpu use steadily increases over time. I have been looking for leaks with memray, but nothing obvious yet.. There is definitely something getting slower over time which makes me think some kind of leak or increasingly inefficient search(?)

the RSS is creeping up ever so slowly - hard to imagine that is the cause of slowness

memray-flamegraph-pub_example.py.47978.html.txt (note: need to download and rename .html)

I wonder if I am not closing my streams cleanly?

                    session._quic.send_stream_data(stream_id, msg.data, end_stream=True)
                    session.transmit()