On Oct 21, 2010, at 21:54 , Steven Timm wrote:

> When Fermilab did our HW evaluation earlier this year we had
> Magny-Cours based machines, 12 core processors, available.  It did
> work but I don't know what if any tunes were done.

OK, I've found the culprit: we use only very little memory, so 16 GB/Node (4 GB/Socket) is quite sufficient. These are actually two DIMMs, using only half of the available channels. This is not a problem in itself, but since the 12-core CPUs are actually pairs of 6-core "nodes", one has to make sure that the memory is properly distributed: once we've populated Slots 1A and 3A (rather than 1A and 2A), performance was back to normal.

That said, there may be a slight difference (3% or so) between SL 5.4 and 5.5, but this would require some more careful checking.


A.

-- 
Ansgar Esztermann
DV-Systemadministration
Max-Planck-Institut für biophysikalische Chemie, Abteilung 105