PGI MPI performance

One of our users has asked a question regarding the PGI MPI setup:

Anyway, I found the latency and bandwith of the GNU MPI library are
both 10 times better than the PGI MPI library.

Can you confirm if either PGI and GNU are installed to perform TCP/IP
between cores or do they both use local memory references.

Can someone tell me which is the case for PGI?
It’s just a default installation - no extra configuration.

Thanks

Hi fbissett3,

The prebuilt MPICH we ship with the compilers is a debug version meant for portability across most any cluster and for use with the PGI debugger. It also uses ethernet so may not be optimized for your interconnect.

For higher performance you may want to build your own MPI libraries targeting your interconnect. OpenMPI is usually a good choice (http://www.pgroup.com/resources/openmpi/openmpi143_pgi2011.htm)

Hope this helps,
Mat