>
>> >Dunno about the cache issues - to be honest I don't know why different
>> >page layouts in physical RAM makes such a big difference...
>> >
> 
>> Maybe TLB misses?
> 
>That's what I thought, but it was pointed out to me that TLB refils
>for x86 are done in hardware and are pretty fast. Upon investigation
>it turned out to be L2 cache misses due to cache line aliasing.
>Note that the cache will cache physical memory, hence the layout of
>physical pages matters much.                                      
>
>The solution: page colouring (thanks to DaveM for his tips). I tried
>that and it works. The patch is crude and can slow down process
>startup times, and it requires a system with plenty of RAM to spare,
>but it does work. If people want to try this, I could be encouraged to
>post a patch.                                                      

Yes I would like to try this on some of our numeric codes. Robert is
quite right -- trying to optimize code the way things are now is
difficult.

Emil

-
Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/
To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]

Reply via email to