Date: Fri, 04 Mar 2005 08:55:02 -0800 From: "Kevin Oberman" <oberman@es.net> To: Nate Lawson <nate@root.org> Cc: current@FreeBSD.ORG Subject: Re: patch: p4tcc and speedstep cpufreq drivers Message-ID: <20050304165502.19EED5D07@ptavv.es.net> In-Reply-To: Your message of "Thu, 03 Mar 2005 21:37:05 PST." <4227F401.7000204@root.org>
next in thread | previous in thread | raw e-mail | index | archive | help
> Date: Thu, 03 Mar 2005 21:37:05 -0800 > From: Nate Lawson <nate@root.org> > > Kevin Oberman wrote: > > OK. This makes me feel a bit better, but I still think I'll leave TCC > > out of the equation as it makes the various frequency steps vary uneven > > to the point that lowering dev.cpu.0.freq would increase performance > > (and the reverse, as well) and it causes my system to hang when > > throttled back too far. It never hangs with TCC disabled although my > > lowest "frequency" is now just 150 MHz. > > Would you test with hint.acpi_throttle.0.disabled="1" instead of > disabling p4tcc? I think p4tcc is not the problem, it's the combination > of the two. I think there are some problems when both the chipset > (externally) and processor (internally) assert STOPCLOCK. If this works > for you with no hangs, I'll commit code to disable acpi_throttle when > p4tcc is present. p4tcc is more efficient than acpi_throttle since the > latter is done through the chipset, giving more chance for race > conditions, latency, etc. Looks like you are right on the button. p4tcc with throttling disabled yields the best results I have seen. The performance is just a little better than the "normalized" value I would expect where throttling produced performance just a little worse. As long as I don't run both, I don't hang at any speed and I don't get increased performance with decreased speed. I really want to try some tests while actively monitoring current draw some day, but it will require hacking on a power brick and I don't have one I can play with at the moment. That would provide some REAL indication of power savings with reduced performance and make tuning more accurate. I am appending the test results. As usual they are on a system running single-user and are very consistently reproduceable with standard variation of less than .4% and usually under .2%. -- R. Kevin Oberman, Network Engineer Energy Sciences Network (ESnet) Ernest O. Lawrence Berkeley National Laboratory (Berkeley Lab) E-mail: oberman@es.net Phone: +1 510 486-8634 1800 81860338 1575 72045833 1350 62122232 1200 57835625 1125 52478629 1050 50840742 900 43377770 750 37008544 675 32357704 600 30049487 450 20917432 300 14160281 225 12938852 150 8663693 150 8660349 225 12941123 300 14179776 450 20935574 600 30005557 675 32258131 750 36811863 900 43108915 1050 50650194 1125 52591176 1200 57684049 1350 62028300 1575 71823521 1800 81454824
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20050304165502.19EED5D07>