+Arjan van de Ven +H. Peter Anvin Some folks internally have been arguing (and with data that appears to support their thesis) that with modern Intel processors, the ondemand CPU governor is actually counterproductive because waking up to decide whether the CPU is idle keeps it from entering the deepest sleep states, and so (somewhat counterintuitively) the performance governor will actually result in the best battery life.
Can you confirm? And if so, what's the first CPU architecture where this is true? Sandy Bridge? Ivy Bridge? And should we perhaps make the on-demand governor automatically deactivate itself on sufficiently new architectures? I looked to see if there was any useful information on www.lesswatts.org, but apparently that domain has been allowed to expire.
Can you confirm? And if so, what's the first CPU architecture where this is true? Sandy Bridge? Ivy Bridge? And should we perhaps make the on-demand governor automatically deactivate itself on sufficiently new architectures? I looked to see if there was any useful information on www.lesswatts.org, but apparently that domain has been allowed to expire.
View 63 previous comments
I tried the latest kernel 3.10 which has Pstate driver for my Ivybridge. I saw my temperatures rise and the frequency even at idle over at 1.80MHz for my i7. I didn't get any noticeable power savings or performance boost with pstate either. I tested this on Manjaro .87.1Sep 18, 2013
I notice poor performance on an E5-2643 in powersave mode. System load is single-threaded processor sending and receiving UDP frames over a VETH pair (same general issue over pair of real igb nics). I get about 400Mbps of 1500 byte pkts tx + tx with powersave, and about 800Mbps with 'performance'. It took a while to find the problem because I had kernel configured to use ondemand by default, and didn't realize on E5 that would translate that to powersave. This is on kernel 3.9.11+.Sep 28, 2013
+Arjan van de Ven +Theodore Ts'o Sorry for jumping in late.. Probably it would be better to initiate such discussions on list, so that people don't miss it. :)
Why do you people say this: "waking up to decide whether the CPU is idle"?
Ofcourse, I am missing some silly stuff.. Work (responsible for timer interrupts) is queued as Deferrable and so it shouldn't wakeup CPU's when they are idle..
Sorry if I was being really stupid :)
Another important update in ondemand governor was done by this patch, which makes it use intermediate frequencies rather than lowest and highest: dfa5bb622555d9da0df21b50f46ebdeef390041bOct 25, 2013
Hi,
I also have a question that someone could answer. What is the support for Intel Haswell processors by both ACPI and P-STATES drivers? It seems that Haswell has an on chip voltage regulator that allows significantly faster DVFS transitions and therefore significant power savings if we can adjust the frequency to the program demands (e.g., memory bound phases of a program could run at lower frequencies). ACPI driver exposes a transition latency of 10usec which is way too coarse grain to be usefull. What about the P-States driver. Does it allow us to have faster transitions in contemporary and future CPU generations? How do you plan to support these features?Nov 21, 2013
+Arjan van de Ven Does NO_HZ_FULL still cause a regression? Also i am on i5 650(clarksdale?) and sadly it uses acpi-cpufreq. Planning on upgrading and wondering if it will enable on new platforms.. I am using 3.14.0Apr 13, 2014
Is P-state driver support on Ubuntu 14.04 (kernel 3.13), I found it'll cause my server hang on during booting.(Xeon E5-2695v2)Nov 21, 2014
Add a comment...