Re: [PATCH v25 0/4] MR3504: ntdll: An implementation of SRWLOCK that closer matches Windows'.

[Yuxuan Shui (＠yshui)]

On Mon Sep 18 09:30:20 2023 +0000, Yuxuan Shui wrote:
> > (I've seen up to 15x slower)
>
> This is because Window's scheduling granularity is larger, so my test
> program ended up sleeping longer. When that's accounted for the overall
> runtime is comparable.
> I also measured the overhead of locking operations themselves. In
> general this MR is comparable with the current implementation, and is a
> bit faster than native. However, when there is no contention (i.e. only
> AcquireShared), native is faster, IIRC it is twice as fast vs the
> current impl, and about 2.5x vs this MR.

OK, I remembered incorrectly, I ran the test again and native is about 1.5x faster, this MR and the current impl is about the same (the measurement's margin of error is a bit large)