| 1 |
On 24 Nov, Stefan G. Weichinger wrote: |
| 2 |
> Stefan G. Weichinger schrieb: |
| 3 |
>> Stefan G. Weichinger schrieb: |
| 4 |
>> |
| 5 |
>>> Since then no crashes, but I would have to test clicking some more stuff |
| 6 |
>>> to really believe ... |
| 7 |
>> |
| 8 |
>> As always, after hitting SEND ... one more crash ... |
| 9 |
> |
| 10 |
> Sometimes it crashes after clicking opera, sometimes after clicking |
| 11 |
> thunderbird, so far never when clicking/starting a gnome-terminal. |
| 12 |
> |
| 13 |
> I am still looking for a pattern or an error-message somewhere ... |
| 14 |
> |
| 15 |
|
| 16 |
This reminds me of a problem we had just recently. |
| 17 |
Have you got a multi-core CPU ? |
| 18 |
If yes, read on. |
| 19 |
|
| 20 |
We have 6 machines here running an identical Gentoo system |
| 21 |
(just different hostname and IP number) |
| 22 |
with a AMD Phenom II quad core CPU and identical mother boards. |
| 23 |
One of them had these random crashes you reported. |
| 24 |
I've totured memory by running up to 3 memtester-processes |
| 25 |
over night - no single fault. Our dealer has replaced the motherboard - |
| 26 |
again no change. Then I suspected the CPU itself although it has stood |
| 27 |
a burnK7 run for several hours. |
| 28 |
|
| 29 |
After the CPU has been replaced the spook has gone. |
| 30 |
I suspect a cache coherence problem. The normal memory tests |
| 31 |
assign a given window of the physical storage to a given core - |
| 32 |
even if run in parallel. But a typical usage under Linux switches |
| 33 |
the core which executes a given thread quite frequently. |
| 34 |
Now the Phenom II has 4 core each with a private 0.5 Mb primary cache |
| 35 |
but a 6 Mb second level cache common to all 4 cores. |
| 36 |
In the BIOS one can opt for all 4 cores using this secondary cache |
| 37 |
or for only a single core using it. |
| 38 |
When a core writes to this cache or to memory all other cores must be |
| 39 |
informed that their private cache is invalid. If this doesn't happen or |
| 40 |
happens a bit too late, a core will fetch invalid (old) memory contents |
| 41 |
which may result in a crash. |
| 42 |
So, if you can, set the BIOS switch that only a single core |
| 43 |
can use the secondary cache. If the problems disappears |
| 44 |
the CPU is broken. |
| 45 |
|
| 46 |
I hope you can solve your problem, |
| 47 |
Helmut. |
| 48 |
|
| 49 |
-- |
| 50 |
Helmut Jarausch |
| 51 |
|
| 52 |
Lehrstuhl fuer Numerische Mathematik |
| 53 |
RWTH - Aachen University |
| 54 |
D 52056 Aachen, Germany |
Archives