[Beowulf] Re: Finally, a solution for the 64 core 4TB RAM market

[Bogdan Costescu]

On Fri, 29 May 2009, Mark Hahn wrote:

> but the point is that 8-socket fat nodes are going to be more 
> expensive; traditionally nonlinearly more expensive.  current 4s 
> boxes are more than 2x 2s cost.

Indeed, but you can use a 4s machine directly, without other 
investment, with best performance for either MPI or OpenMP job, while 
the 2 machines with 2s will need a fast interconnect to achieve a 
similar setup and performance, adding to the cost of the nodes.

If the application doesn't scale too well to high number of nodes 
anyway (f.e. a molecular dynamics code with a small size molecular 
system, using PME, etc.), nodes which are fat enough to accomodate it 
will offer most often the best performance and will be much easier to 
manage - no need for extra drivers, testing of cables, routing/mapper, 
etc. And 32-64 cores per node would be exactly the sweet spot for the 
types of jobs that my colleagues are running...

-- 
Bogdan Costescu

IWR, University of Heidelberg, INF 368, D-69120 Heidelberg, Germany
Phone: +49 6221 54 8240, Fax: +49 6221 54 8850
E-mail: bogdan.costescu at iwr.uni-heidelberg.de