At this point, two environment variables were then tested, namely MPICH_PTL_EAGER_LONG and MPICH_PTL_MATCH_OFF.
The environment variable MPICH_PTL_EAGER_LONG=1, uses the EAGER path for all messages, regardless of their size.1
The environment variable MPICH_PTL_MATCH_OFF=1 is useful for simulations which have many small messages and is latency bound.
The `200.40' experiment (2003 case running for 40 time steps) was run three more times, once with MPICH_PTL_MATCH_OFF set, once with MPICH_PTL_EAGER_LONG, and once with both environment variables set.
The actual timing result are not contained in this report; however, it was found that neither environment variables had an adverse effect on performance, irrespective of the process count. For large processor counts, it was found that setting MPICH_PTL_EAGER_LONG could give up to a 7% speed-up, especially around the optimum number of processors. Thus, we employed MPICH_PTL_EAGER_LONG=1 for all other benchmarking in this report, and include it in the example batch scripts4.2.
Both the 1003 and the 2003 benchmark cases are available on the HECToR's OpenFOAM web page .