Charm++ parallel programming system

[Fouzhan Hosseini <F.Hosseini AT leeds.ac.uk>]

I got a bounce email for charm AT lists.illinois.edu, so I am just re-posting here. Sorry, If it is posted multiple times!

Hi all,

Happy new Year!

I have developed a Charm++ application, and tested its performance on our cluster using both an MPI and ibverbs Charm++ build.

The MPI build is twice slower than ibverbs build. I know ibverbs build should be faster but I was surprised to see it twice faster than MPI build.

Timeline profiling looks very similar for both cases, except one is taking much longer. I also notice that MPI build has a much higher memory usage.

I was wondering whether it is generally expected to see such difference in performance between MPI and ibverbs build or I am using MPI build in a wrong way.

Would you be so kind to share any link/document that explain communication level of charm and its integration with ibverbs/MPI, if there is any? I am wondering why  MPI build has a higher memory usage?

Our cluster has a Mellanox interconnect.

Many thanks,

Dr Fouzhan Hosseini
Research Fellow
School of Computing
University of Leeds
Leeds, LS2 9JT