Re: [gentoo-user] CPU "choking" under high load

[Paul Hartman <paul.hartman+gentoo@gmail.com>]

On Fri, Mar 26, 2010 at 3:47 PM, Daniel Quinn
<expendable.0@danielquinn.org> wrote:
> I don't know if this is a hardware issue or not, but I thought that maybe I'd
> configured my kernel incorrectly and that this might be a known issue someone
> here has run across in the past so here goes:
>
> My computer is a pretty impressive AMD 64 X2 Dual Core Processor 6400+ box
> with 2GB of RAM and for the most part, I just use it to write code at work.
> However, whenever I'm doing something CPU-intensive, two things happen:
>
>  * The load on the box goes up to 4
>  * The box intermittently wobbles from running at full-speed and dropping to
>    a crawl.  This is best seen while watching Flash videos online,
>    compressing/encoding video, or compiling.  Everything is fine for a few
>    minutes, then suddenly the rate of compiling/compression/playback etc.
>    drops to a crawl for about 1-3minutes, then back up to full speed.
>
> I don't know why it's happening.  I've tried various kernel options with no
> change in behaviour.  Outside of that though, I don't know what to try.
> Suggestions welcome :-(

My initial thoughts, it sounds like what I experienced a few kernels
ago when they introduced the new group scheduler features. In my case
I disabled "Group CPU scheduler" in kernel config and changed from
SLUB back to SLAB. Also perhaps try to change your preemption model,
timer frequency, NO_HZ mode. These things have all had noticeable
differences to me in responsiveness over the years.

If it's not a kernel issue, perhaps use "nice" to tame CPU-hungry
processes, or "ionice" if they are disk-intensive.

Be sure your Video Card is not overheating or going into low-power
mode. Nvidia cards go into a super-slow-mode when they overheat
(happened to me 3 times when the fan on the card died). When that
happened, for example my FPS in glxgears would go from thousands down
to double digits... I could see lines painting on the screen, it was
painfully slow.

Also, if your CPU has frequency scaling make sure it's not gone crazy,
cat /proc/cpuinfo to see CPU speed, or powertop might be a more
friendly way to view the P-States.

Download memtest86+ from memtest.org and run it to be sure you don't
have any faulty RAM (in my experience tests 5 and 8 are the only ones
that ever show any errors, so you can save time by only running one
full pass of all tests and then a multiple passes of 5 and 8 just to
be sure it's good)