This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [patch v1] nptl: smarter not-parallel-ing
- From: Florian Weimer <fweimer at redhat dot com>
- To: DJ Delorie <dj at redhat dot com>
- Cc: libc-alpha at sourceware dot org
- Date: Wed, 14 Aug 2019 14:41:35 +0200
- Subject: Re: [patch v1] nptl: smarter not-parallel-ing
- References: <xnk1bgoi3e.fsf@greed.delorie.com>
I did some benchmarks with this patch.
* Cold case.
Newly build source tree, “make subdirs=login check” (to populate the
testroot), followed by “make subdirs=nptl check”, timings for the
latter. All tests were for x86-64, timings in seconds (wall clock).
Appropriate -j options supplied.
Before:
8-thread system: 399 399 397 400 398 400 400
256-thread system: 392 389 389 391 394 393 392
After:
8-thread system: 372 375 373 376 373 374 375
256-thread system: 357 357 356 361 358 358 357
* Warm case (incremental make).
Timings of “make subdirs=nptl check” after the first test suite run.
Appropriate -j options supplied.
After:
8-thread system: 1.541 1.014 1.025 1.016 1.030 1.029
256-thread system: 1.789 1.720 1.142 1.119 1.133 1.143
(I didn't bother to time the before case, considering these low
numbers.)
So it seems to me that performance-wise, this is a small overall win.
Thanks,
Florian