wamr jit/aot faster than native · wamr

Stream: wamr

Topic: wamr jit/aot faster than native

Mats Brorsson (Dec 12 2023 at 14:54):

Under what circumstances will a WebAssembly program execute faster with iwasm using the llvm jit? I have made a study and there is one program, PolyBench/correleation which can be found in this repo: https://github.com/matsbror/wabench/tree/ace5g, which on x86_64 executes faster with iwasm/llmv-jit than natively.

It seems conter-intuitive. I have tried to isolate the function in question as a WebAssembly module of its own and disassemble it, but, unfortunately, then the performance advantage disappeared.

Any ideas?

GitHub - matsbror/wabench at ace5g

Contribute to matsbror/wabench development by creating an account on GitHub.

Jämes Ménétrey (Dec 12 2023 at 15:49):

Hello,

Check the following studies for similar observations:

Twine: An Embedded Trusted Runtime for WebAssembly (https://arxiv.org/abs/2103.15860), Section V.B.
Not So Fast: Analyzing the Performance of WebAssembly vs. Native Code (https://www.usenix.org/conference/atc19/presentation/jangda), Section 6.3.

Basically, we concluded that a slightly faster execution using AOT may be due to fewer L1 instructions cache misses.

Cheers

Last updated: Oct 23 2024 at 20:03 UTC