Performance regression in 1.6.0.0

[NathanRichard]

Hello,

I saw that some performance work has been done in the 1.6.0.0 release so I conducted my own benchmarks using the python interface.
According to my benchmark, the 1.6.0.0 actually brought a reproducible performance regression.

Here are my results (on a moderately modern Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz CPU):

Test 20 Mio Binaries	Old Zfec (1.5.7.4)	New Zfec (1.6.0.0)
Sequential k=10 m=30	86.83 Mio/s	59.56 Mio/s
Parallel (Threadpool with 16 workers) k=10 m=30	377.74 Mio/s	285.68 Mio/s
Parallel (Processpool with 16 workers) k=10 m=30	122.19 Mio/s	113.54 Mio/s

I can reproduce these issues between runs and on a variety of laptop, desktop and server x86 CPUs. My tests were conducted under ubuntu 20.04 and ubuntu 22.04 with python 3.10 and python 3.12.

I have not tested the release commit by commit so I cannot pinpoint the issue. However, the closest change seems to be #114.

This MR claims to have doubled the compute speed of zfec, do you have more information about how these tests were conducted and how we could explain these discrepancies ?

Best regards,

[hacklschorsch]

Tagging @itamarst

[hacklschorsch]

If you haven't come accross them @NathanRichard , #102 has some performance numbers.

IDK if we shouldn't revert that change. It also made Tahoe-LAFS less compatible for older HW.

BTW some wilder performance enhancements can be found in zfex - would be interested if you can replicate their benchmarks!

[NathanRichard]

Hello,

Sorry for the delay of my reply.

@hacklschorsch Thank you for the link to the performance numbers, however I do not understand with which test they were produced (and hence with which parameters k/m). Was it with bench_zfec ?

I tried benchmarking zfex in 2023 but at the time the speedup where not available through the python interface.

If you want, I can provide my benchmarking script to see if you can reproduce the regression yourself.

Best regards