MPI on dual core processors?

[f95toli]

Does anyone know if there are any plans for a MPI imlementation that supports dual-core processors?
And if so, will the two cores "behave" as two processors (from the MPI point of view that is, i.e. can I run one process on each core)?

Or maybe I should start by asking if such an implementation is even possible?

 

[Armitage]

I would guess that MPI won't care - the OS takes care of scheduling processes & threads across CPUs and/or cores on CPUs. MPI operates well above this level, and should work just as it does on multi-cpu machines.

 

[CTho9305]

Given that the dual-core CPUs have shared memory, if you could, it would probably be most efficient to count each package (rather than each core) as a node so you don't pay the overhead for passing messages when you could just use the shared memory system.

 

[f95toli]

Originally posted by: CTho9305
Given that the dual-core CPUs have shared memory, if you could, it would probably be most efficient to count each package (rather than each core) as a node so you don't pay the overhead for passing messages when you could just use the shared memory system.

But that requires an implementation that "understands" that a dual core processors can be efficiently used with two processes, otherwise there is no way to use the optimal topology in the program.

The cluster I use is based on a network of quad-opterons (mpich/Linux) and I can set both the number of nodes and CPUs when I schedule a job; if there is a lot of communication overhead the configuration can have a huge impact on the speed.

 

[Armitage]

Originally posted by: CTho9305
Given that the dual-core CPUs have shared memory, if you could, it would probably be most efficient to count each package (rather than each core) as a node so you don't pay the overhead for passing messages when you could just use the shared memory system.

I haven't followed the dual-core news much - I expect processor affinity will become even more important. But again, I suspect it will happen at the OS level.

Probably your best bet would be to multithread the jobs on each node to take advantage of the multi-core architecture, then just start one job on each node. So within a node you can use shared memory, and only need message passing between nodes.

That's what I'm doing on our dual-xeon cluster - though I actually start two jobs/node because some non-threadsafe code I have to use bottlenecks it otherwise.