2.6.10-rc2: strange behavior on dual Opteron w/ NUMA

From: Rafael J. Wysocki
Date: Mon Nov 15 2004 - 17:11:18 EST


I am observing a strange behavior of the 2.6.10-rc1-mm5 and 2.6.10-rc2 kernels
on a dual-Opteron box with NUMA (Tyan Thunder K8W). It occurs in specific
conditions, but seems to be 100% reproducible.

The situation is the following: I run X with KDE 3.3.1 (SuSE binaries) on SuSE
9.1 (x86-64), I run gkrellm in the background, I edit a text file using Kate,
there's a Konqueror window open and at the same time I play OGG files using
Noatun (not so unusual, I'd guess). Then, after some time (variable), Kate
suddenly starts to repeat the most recently typed character (I have to leave
its window to stop this), Noatun stops playing and arts reports CPU overload.
gkrellm also shows the sudden increase of CPU load (apparently, only one CPU
is almost 100% loaded and the load sometimes "flows" from one CPU to the
other). It all takes about 10 seconds. Then it calms down, and I can start
playing music and typing again. Sometimes Kate behaves normally, but every
time it happens a CPU gets overloaded which causes Noatun to stop playing and
CPU overload is reported (in such cases, the CPU is overloaded with _system_
load). So far, it's happened every time I ran the box with one of the said
kernels, and it can happen several times in a row (at random times).

It doesn't occur on the 2.6.10-rc1 kernel (for sure) and, AFAICT, it doesn't
occur on the previous -mm kernels (I must admit I haven't tested them
thoroughly on the dual box, though). I also do not observe this kind of
problems on a single-CPU box with a similar setup, but I usually don't use
Noatun on it.

I suspect that this has been intorduced in 2.6.10-rc1-mm5, so if you have any
ideas, please let me know. If you need more information, please let me know


- Would you tell me, please, which way I ought to go from here?
- That depends a good deal on where you want to get to.
-- Lewis Carroll "Alice's Adventures in Wonderland"
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/