Stream: git-wasmtime

Topic: wasmtime / issue #5004 x86_64: align loop headers to 64 b...

Wasmtime GitHub notifications bot (Oct 04 2022 at 18:17):

cfallin commented on issue #5004:

@pepyakin , thanks for this change. However, I'm not sure that we want to take it as-is; or at least, I'd like to see more data.

The Sightglass runs you provide are all over the place -- in some cases this is better by 5-10%, in other cases main is better by 5-10%. There doesn't appear to be a clear trend to me.

What's more concerning, the instantiation times have wild swings as well, but this shouldn't have been affected at all by this change. This makes me concerned about the reliability of the measurement setup and confidence-interval computation.

Would you be willing to take a few of the larger benchmarks (spidermonkey and bz2, say) and run them by precompiling to .cwasms then measuring time with your tool of choice (I like hyperfine) for a wasmtime run ... command? I'm curious how much variance this sees and what trends we'll get.

I share @bjorn3's concern about code size here as well. Aligning every loop to a 64-byte granularity is likely to bloat some cases nontrivially. Could you report the effect on code-size (the .text segment of the .cwasms) for benchmarks as well?

Wasmtime GitHub notifications bot (Oct 11 2022 at 14:49):

pepyakin commented on issue #5004:

That's a good catch. I must admit that I did not notice that it was about instantiation! Probably something was running on my test machine and ruined the measurement. I also agree with the points made by @bjorn3. In retrospect, that seems obvious :face_palm:

I had an impression that this PR would be a quick one, but it seems I was wrong = ) I will try to further it, but that won't be my primary focus, so it will likely take some time[^1].

[^1]: Should anyone feel the acute need/or just want to take over for other reasons, feel free to take it on.

Last updated: Nov 22 2024 at 17:03 UTC