On Wed, 24 Sep 2014, Rafael Vega wrote: It sounds like there are two things that might work... use an older kernel (Is there something in 3.14 you must have?, maybe 3.16 will fix if you wait.) Set the default governor to performance at compile time (or maybe boot time). Personally, under stress test, I have found that with all cores at 100% Performance runs cooler than Ondemand or userspace with some cores set lower speed. The only place I find a difference is at idle performance seems a bit warmer... but at idle the temperature is way lower anyway. Is boost turned off in bios? The new kernel needs to have speed step on to use Intel's boost technology. Boost over speeds the cpu till it over heats then backs off... over simplified yes. Performance may put the chip into boost mode (performance is the highest speed after all) all the time. Can you get it into userspace? If so you could set the speed manually. I just checked. Boost speed for your chip is quite high(3.4Gh) I only have an i5 with 4 cores no HT. (maybe I shouldn't say only as yours has only two cores 4 with HT) With the cores set performance and stress set to 100% on all cores, I am still 10deg less(~60C) than Max running temp. (70C) Temperature shutdown should be a lot higher than that (over 100C ... both your chip and mine)). Are you using a temperature monitor? The temperature should settle after about 10min at a stress level. (both my tests and Intel specs agree on the 10 min thing) I am using 3.13 lowlatency, but 100% stress should do the same either way. It sounds like you machine is running really hot... or is running over 2.7G I have been using psensor to monitor temperature. Mine is not a laptop though. Just a note, any testing I have done has shown that running my 3Gh + (without boost) CPU at 800Mh (with the userspace governor) gives better lowlatency than ondemand at full speed does. Hyperthreading and ondemand: Ondemand affects even higher latency. HT only seems to be an issue for latency below jack at 64/2, at least in my tests. -- Len Ovens www.ovenwerks.net

On Thu, Sep 25, 2014 at 4:13 AM, Rafael Vega &lt;email.rafa(a)gmail.com&gt; wrote: I've been noticing the issue of not being able to change frequency governer on -rt kernels here. I've posted to the -rt mailing list, details here (click "next in thread" to follow convo): http://marc.info/?l=linux-rt-users&m=140897493204106&w=2 In short, there seems to be something going wrong with locking / unlocking an RW lock that is used during a CPU freq scaling. I have pin-pointed the issue to a pretty specific place in <kernel>/drivers/cpufreq/cpufreq.c but I don't know how to fix it. Note that I'm running 3.14.12-rt9-1-rt (Arch linux, from AUR) but the issue seems identical to me.. I have an older CPU (Core2Duo from about 7 years ago), but I read in the forum link you have an i7? That shouldn't have the same issue... I'll post to the -RT list, mentioning your email, and you system config, and see if the devs there think they're releated or different. Cheers, -Harry -- www.openavproductions.com

On Wed, 24 Sep 2014 22:13:48 -0500 Rafael Vega &lt;email.rafa(a)gmail.com&gt; wrote: I can confirm both issues on the archlinux 3.14-rt kernel. Maybe best to seek a resolution on the linux-rt-users mailing list. If you have a recent Intel cpu, you might be able to work around the problems using the /dev/cpu_dma_latency and /sys/devices/system/cpu/intel_pstate interfaces. -- Joakim