From mboxrd@z Thu Jan 1 00:00:00 1970 From: Mark Bidewell Subject: Re: Possible CPUFreq governor Date: Mon, 02 May 2005 08:51:25 -0400 Message-ID: <4276224D.5020005@alumni.clemson.edu> References: <53866.130.127.49.120.1114558769.squirrel@130.127.49.120> <20050427105350.GX2298@poupinou.org> <46736.141.228.156.225.1114600117.squirrel@www.difo.com> <426F85F5.7010700@alumni.clemson.edu> <20050427140711.GA2298@poupinou.org> <426FBBB1.7000001@alumni.clemson.edu> <20050427173828.GC2298@poupinou.org> <46372.130.127.49.120.1114625907.squirrel@130.127.49.120> <20050427190416.GE2298@poupinou.org> Mime-Version: 1.0 Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20050427190416.GE2298@poupinou.org> List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: cpufreq-bounces@lists.linux.org.uk Errors-To: cpufreq-bounces+glkc-cpufreq=m.gmane.org@lists.linux.org.uk Content-Type: text/plain; charset="us-ascii"; format="flowed" To: Bruno Ducrot Cc: cpufreq@lists.linux.org.uk I have been examining the structure and design of the ondemand govenor further. Would it be accurate to characterize your concern as basically that under heavy process load, the p-state switching latency becomes longer than the timeslices and thus could dominate the CPU? Or is there an issue with processor damage? Bruno Ducrot wrote: >On Wed, Apr 27, 2005 at 02:18:27PM -0400, Mark Bidewell wrote: > > >>It might help to explain how the governor is used. I have a user-mode >>daemon which samples the ACPI temperature every 2 seconds and communicates >>the speed at which non-interactive processes should be run via the sysfs >>interface. The scheduler will then determine a processes interactivity >>and change the CPU speed accordingly. In theory the processor could >>change speeds every process switch. >> >> > >That why I have to do tests in case there is a lot of those process >switch but when modifiying max frequency instead. > > > >>This governor is indeed similar to the ondemand governor. The distinction >>is that while the ondemand governor expands CPU performance as load >>increases, tempscale must limit performance. In my opinion, the ultimate >>solution would be a combination of the two governors. That is, a governor >>which would react in the following way: >> >>1) If the processor is idle, run at low speeds to conserve power (ondemand). >>2) Increase performance with demand as long a temperature is low (ondemand) >>3) If temperatures get to high, throttle compute-bound processes (tempscale) >> >>I have done some limited experiments along this line using a user-mode >>daemon to switch governors (as well as speeds) between tempscale and >>ondemand as temperatures change. >> >> > >That's why I suggest to touch max frequency, not the actual frequency. > >This must be done outside a governor, so that you can have both actually. >I'll send you a patch after writing it for testing purpose so that you will >see what I have in mind. > > >