On Wed, 12 Mar 2008, Rob Malpass wrote:
> My company now think it might be better to buy a higher spec machine
> and run the cluster as 4-6 virtual machines.
If you have the source code and can make it work on a current Linux
distribution, the best is probably to buy a multi-core and/or
multi-processor machine and rebuild your application with the latest
LAM/MPI version or probably even better OpenMPI. This way the
application should not see any difference in the way it is run - it's
still MPI, except that all ranks will be running on the same machine -
and communication through shared memory is faster than through the
network, be it physical or virtual. The setup is very simple, not
virtualization is involved...
--
Bogdan Costescu
IWR, University of Heidelberg, INF 368, D-69120 Heidelberg, Germany
Phone: +49 6221 54 8869/8240, Fax: +49 6221 54 8868/8850
E-mail: bogdan.costescu_at_[hidden]
|
