[molpro-user] HP-MPI MPI_Send out of resources

Mikael Johansson mikael.johansson at helsinki.fi
Tue Jun 12 21:34:17 BST 2007

Hello again!

On Sun, 10 Jun 2007, Mikael Johansson wrote:

> I'm running Molpro 2006.1 pl 96 on a HP-MPI cluster (with LSF-HPC/SLURM), and 
> keep getting the following type of error messages with slightly larger 
> CCSD(T) jobs (smaller jobs run fine):
> molprop_2006_1_i8_x86_64_mpi.exe: Rank 0:1: MPI_Send: VAPI_register_mr() 
> failed: addr 0x2b332a5f20, len 65536: Out of resources

No real solution yet, but forcing the transfer protocol to TCP makes the 
calculations proceed. Don't know what kind of performance hit this implies 
(if any) though. For the record, also tried uDAPL, which caused a similar 
MPI_Send error to the above. So for example:

   export MPI_IC_ORDER=TCP
   export MPIRUN_SYSTEM_OPTIONS="-subnet x.y.z.å"

Apparently a HP-MPI ( problem, on another machine with MVAPICH 
InfiniBand, the problem doesn't occur.

Have a nice day,
     Mikael J.

More information about the Molpro-user mailing list