Hi.
Some update on the issue after quite some time. At the moment, the only config
that seems to work somewhat decently is using the cpuidle patches for the
kernel without any governors. As I said earlier, powertop then gives a message
that the cpu is 100% busy and no c-states are entered. The machine is not at
its coolest best, but is somewhere in between good and bad states. If I use
ladder governor, it consumes about 40W, and in bad state, it consumes about
60W. This config consumes about 50W. Aside, giving idle=poll boot parameter
sets my cpu on fire (as is known).
A few more observations. Running a non smp kernel seems to fix the issue.
Frankly, I am fine with the performance of a non smp kernel. There is
absolutely no difference in single benchmarks and games (as expected). But it
really shows up when I run parallel heavy tasks :). So I might even go on to
say that its a smp issue. As such, it pretty much looks like a race-condition.
I tried to track the problem if any specific kernel module causes this, by
compiling as many kernel features modular as possible. Without the cpuidle
patch, as soon as I insert the module acpi-processor, I get the messages :
Marking TSC unstable due to: possible TSC halt in C2.
Time: hpet clocksource has been installed.
Of course, acpi c-states dont show up in powertop before I use this module.
When using cpu-idle patches, this message shows up when I insert the ladder
module. When using the menu governor module in cpu-idle patches, the c-states
do show up and my cpu stays in C0 100% of the time.
When it enters into bad state, even removing and re-inserting the
acpi-processor module doesn't help. However, if I hibernate and turn on again
(I am using suspend2), it returns to good state.
Aside, there are two cpufreq control drivers for my centrino cpu - the
acpi-cpufreq driver and the speedstep-centrino driver. If I don't use any
cpufreq driver, the performance is horrible and halved, even though my cpu runs
at full 1.6GHz. Every benchmark result is halved. Wonder how so. Meanwhile,
when using cpufreq control, the speedstep driver gives a marginally better
performance than the acpi-cpufreq (say 29 seconds instead of 35 seconds in
superpi, and 61fps instead of 58 fps in doom3.)
Anyway, as of now, I am happy with the cpuidle patches with no governor.
Although I would love to have the ladder governor working to stop my fans from
whining frequently. Of course, I do insert the ladder module when I am on
battery. Although the non smp kernel works, I am too greedy not to use it ;-).
So for the moment, somewhat content :).
Hirakendu.
hirakendu das <[EMAIL PROTECTED]> wrote: This time around I am running a kernel
with cpu-idle patches. Powertop now says :
< CPU was 100% busy; no C-states were entered >
and wakeups-from-idle per sec are about 100. The battery consumption is just a
little worse from the good state, but way better from the bad state. Can
somebody tell me if this is good or bad - i.e., is my cpu really busy or not ?
(It doesn't seem so to me though :p.)
Thanks,
Hirakendu.
---------------------------------
Pinpoint customers who are looking for what you sell.
_______________________________________________
Power mailing list
[email protected]
http://www.bughost.org/mailman/listinfo/power
---------------------------------
Moody friends. Drama queens. Your life? Nope! - their life, your story.
Play Sims Stories at Yahoo! Games. _______________________________________________
Power mailing list
[email protected]
http://www.bughost.org/mailman/listinfo/power
