[Condor-users] MPI and multiple SMP machines

[Jens Harting <jens@xxxxxxxxxxxxxxxxxxxx>]

Hi,

our Condor pool consists of a number of single and dual CPU Linux machines 
and we often like to submit MPI jobs using 2 CPUs. I.e., they would nicely 
fit onto a single dual CPU SMP machine. However, since all dual CPU PCs 
are configured as dedicated schdulers, Condor does spread the individual 
MPI processes between different machines. Usually, one process ends 
up on PC A, while the other one is on PC B. Since we only have a 100MBit 
network, the interconnect between the CPUs is too slow. It would be great 
to tell Condor to start as many MPI processes as possible on a single 
machine. So far, we do that by giving the machine name as a requirement in 
the submission script. However, this is not efficient at all if there are 
many SMP machines in the pool.

Is there any way to tell Condor to handle MPI jobs more efficiently?

Thanks for your help,

Jens