search

ptkila / cowichan

Automatically exported from code.google.com/p/cowichan
0 stars 2 forks source link

Improve MPI performance #6

Open GoogleCodeExporter opened 9 years ago

GoogleCodeExporter commented 9 years ago 
Boost MPI has been used with:
- MPICH on windows/linux (1 node, 4 cpus, 4gb ram; 20 nodes, 2 cpus, 2gb 
ram, 100Mbit lan)
- MPICH2 on windows (1 node, 4 cpus, 4gb ram)
- OpenMPI on linux (168 nodes, 8 cpus each, 16gb ram, InfiniBand)

The performance results are not impressive. None of the problems scaled 
beyond 7x. About a third of problems didn't scale.

Some possible reasons:
1. Boost MPI (1.37) and MPICH2 (1.1.1) don't work well together (MPICH 
scales much better in our experiments!)
2. MPICH is old/abandoned.
3. Boost MPI interface uses type 'int' for array sizes for transfering 
data, which is defined as a 32-bit integer by most compilers. Transfering 
data > 4gb is problematic on 64-bit clusters.

Original issue reported on code.google.com by andrew.b...@gmail.com on 16 Aug 2009 at 2:22

GoogleCodeExporter commented 9 years ago

Original comment by andrew.b...@gmail.com on 16 Aug 2009 at 2:23