On Tue, 13 May 2008, Markus Schmidberger wrote:
> Can we declare each processor as a node?
What's the point ? You can ask LAM/MPI to assign more than one process
per node with the 'cpu=X' syntax.
> The 4 processors (now nodes) at one computer have to share the main
> memory. There could be package collisions or latency ?
They have to share the main memory either way. As soon as a job has
started a second, third, etc. process on the node, they will all share
the main memory.
If LAM/MPI is compiled with shared memory support, it will use it for
communication between processes on the same computer. If LAM/MPI is
compiled without shared memory support, it will use TCP/IP for
communication between processes on the same computer - which is
usually inefficient. Shared memory is considered low latency and there
are no packets to produce collisions; TCP/IP over localhost is also
usually without collisions.
> Is there any literature concerning this problem?
Yes. The LAM/MPI web page contains a FAQ for a start.
--
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]
|
