Date: Wed, 25 May 2011 23:57:03 +0200 From: "Ronald Klop" <ronald-freebsd8@klop.yi.org> To: "FreeBSD Stable Mailing List" <freebsd-stable@freebsd.org>, "Josh Carroll" <josh.carroll@gmail.com> Subject: Re: [poll] hyperthreading_allowed, hlt_logical_cpus, mp_watchdog Message-ID: <op.vv14etsc8527sy@pinky> In-Reply-To: <BANLkTik8G4vYVFGwB%2BS-HtBmnas_TgY-pg@mail.gmail.com> References: <4DDBB0C4.5010109@FreeBSD.org> <irgcl2$91e$1@dough.gmane.org> <BANLkTik8G4vYVFGwB%2BS-HtBmnas_TgY-pg@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, 24 May 2011 18:27:47 +0200, Josh Carroll <josh.carroll@gmail.com> wrote: >> Whatever you do, please leave at least some way (at least a tunable) to >> enable/disable HTT - some workloads are better with, and some without >> it, >> and some BIOSes are unreliable in enabling/disabling it :) > > I noticed that disabling HTT via the machdep.hyperthreading_allowed > tunable resulted in worse performance than disabling HTT via the BIOS. > I'm not sure if that is to be expected due to how the scheduler > ignores those cores or what, but just an observation I found while > testing my i7 2600k to see if I should leave HTT on or off for my > particular workloads. > > I was expecting only certain workloads to perform better with HTT on > and others better with it off. However, I haven't found a task that > performed better with HTT off. Some are almost equal (within a margin > of error), but most are faster with HTT enabled (see below for the > testing I did). > > For the tests below, I used an mdmfs for /usr/obj for all > buildworld/buildkernel tests and they were all performed after caching > the contents of /usr/src. I also ran each test 5 times and for each > data point, I took the best run time. For the ffmpeg tests, I used > -threads 0 to let it pick the number of threads to use. It seems > hyperthreading on the Sandy Bridge i7 may be worth leaving enabled, > though of course what I've tested below is just a small sample of > possible tasks. So as with any benchmark, the only relevant tests are > the things you use your machine for :) My workloads happen to all > benefit from having HTT enabled. > > Josh > > > buildworld (w/o HTT=-j5, w/HTT=-j11) > =============================== > HTT off: 558.65 seconds > HTT on: 501.01 seconds > > buildkernel (w/o HTT=-j5, w/HTT=-j7) > =============================== > HTT off: 188.80 seconds > HTT on: 187.03 seconds > > ffmpeg (10 min SD input - xvid/mp3) > =============================== > HTT off: 7.75 seconds > HTT on: 6.28 seconds > > ffmpeg (5 min 720p input - h264/aac) > =============================== > HTT off: 20.08 seconds > HTT on: 12.97 seconds > > ffmpeg (1 min 1080p input - h264/aac) > =============================== > HTT off: 38.91 seconds > HTT on: 30.79 seconds > _______________________________________________ > freebsd-stable@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-stable > To unsubscribe, send any mail to "freebsd-stable-unsubscribe@freebsd.org" I ones read about progress Intel made with HTT in newer cpu's. So it might very well depend on cpu type if the rumors about worse performance with HTT are true or not. Ronald.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?op.vv14etsc8527sy>