Messages in this thread | | | From | Linus Torvalds <> | Date | Fri, 17 Nov 2023 13:57:32 -0800 | Subject | Re: [linus:master] [iov_iter] c9eec08bac: vm-scalability.throughput -16.9% regression |
| |
On Fri, 17 Nov 2023 at 11:13, Borislav Petkov <bp@alien8.de> wrote: > > I wouldn't want to optimize some weird loads. Especially if you have > weird loads which perform differently depending on what uarch > "optimizations" they sport. > > I guess optimizing for the majority of machines - modern FSRM ones which > can do "rep; movsb" just fine - is one way to put it. And the rest is > best effort.
Yeah, we shouldn't optimize for microbenchmarks in particular.
The kernel robot performance reports have been interesting, because they do end up often pointing to real issues. But we've had these kinds of things too, where the benchmark is just odd and clearly happens to trigger something that is just very machine-specific.
So I don't think we should use either of these benchmarks as a "we need to optimize for *this*", but it is another example of how much memcpy() does matter. Even if the end result is then "but different microarchitectrues react so differently that we can't please everybody".
Linus
| |