Charm++ parallel programming system

[Marius Micluta <marius AT biochim.ro>]

Hello,

I had a similar problem about a year ago on a cluster of the same size (Bull 
NovaScale), but with Gigabit Ethernet instead of IB, and the solution posted 
then on this list worked like a charm for me.

I always use the '++skip_cpu_topulogy' switch when launching a charm++ job. By 
that time it was undocumented and I had to apply a patch from the CVS, but now 
it should be available in the production release.

Topologies are different on Ethernet and IB, but for a small cluster it should 
matter less (and a less optimally tuned run is however better than a freeze), 
so you could give a try to the switch above.


Hope it helps,

Marius Micluta

--

Marius Micluta

Institute of Biochemistry of the Romanian Academy
296 Splaiul Independentei
060031 Bucharest
Romania
Tel: +40 21 223 90 69
Fax: +40 21 223 90 68