Re: [v2] New benchtest: pthread locks

[Siddhesh Poyarekar <siddhesh@gotplt.org>]

On 10/10/20 9:25 AM, DJ Delorie via Libc-alpha wrote:
> 
> v2 changes do_bench_2 to run multiple passes and provide a mean and
> standard deviation as well as some other statistics, since I got tired
> of running the whole benchmark multiple times and computing them
> manually ;-) 
> 
> From 94653e93eec9403ebf61b00179799751ce12088d Mon Sep 17 00:00:00 2001
> From: DJ Delorie <dj@redhat.com>
> Date: Wed, 7 Oct 2020 17:04:12 -0400
> Subject: New benchtest: pthread locks
> 
> Performance benchmarks for various posix locks: mutex, rwlock,
> spinlock, condvar, and semaphore.  Each test is performed with
> an empty loop body or with a computationally "interesting" (i.e.
> difficult to optimize away, and used just to allow lock code to
> be "hidden" in the filler's CPU cycles).

This is good to commit, thanks for writing these tests!

> +/* The point of this benchmark is to measure the overhead of an empty
> +   critical section or a small critical section.  This is never going
> +   to be indicative of real application performance.  Instead we are
> +   trying to benchmark the effects of the compiler and the runtime
> +   coupled with a particular set of hardware atomic operations.
> +   The numbers from this benchmark should be taken with a massive gain
> +   of salt and viewed through the eyes of expert reviewers.  */

That is a great comment.  All benchmarks should have a goal set like
this so that it is clear to whoever evaluates them that just getting
better numbers in a specific benchmark is not good enough and one would
have to evaluate why a certain set of results supports their optimisation.

Siddhesh