Hi Terri,

Here is an my recommendation regarding the number of MPI processes,
OpenMP threads, and nodes when running the hybrid version of MrBayes
3.1.2 on Lonestar, Abe, or Queen Bee.

Obtain the number of runs (nruns), chains (nchains), and characters
(nchar) from the input data, or ask the user to input those
separately.

Determine whether the data set is partitioned, either from the input
data or via separate query.

Check that nruns * nchains is a multiple of

   2 on Lonestar or
   4 on Abe and Queen Bee.

If not, ask the user to change parameters to ensure that.  If this
test makes the interface overly complicated, just check that nruns *
nchains is a multiple of 4 regardless of computer.

In all cases, the number of MPI processes is

   nruns * nchains.

The number of OpenMP threads is

   2 if the data set is partitioned,
   4 if the data set is not partitioned and nchar < 2,000,
   4 if the data set is not partitioned, nchar >= 2,000, & running on Lonestar,
   8 if the data set is not partitioned, nchar >= 2,000, & running on Abe or QB.

The number of nodes is then

   MPI processes * OpenMP threads / cores-per-node, where

cores-per-node is 4 for Lonestar and 8 for Abe or Queen Bee.

All runs will use the hybrid MPI/OpenMP executable.

The hybrid code is on Abe at

   /u/ncsa/pfeiffer/MrBayes-3.1.2h .

The makefile is

   Makefile.icc.hybrid .

Sample input and output files are at

   /u/ncsa/pfeiffer/MrBayes-3.1.2h_abe/150 .

These were all run with MVAPICH2.  To run with Open MPI, the run
scripts need to be changed similarly to those for RAxML.

* Let me know if you have any questions.

Thanks,  Wayne