Hello, I'm trying to run SIESTA on two 8-proc nodes.

I use MPICH2 1.3.1 and hydra (default) process manager

When I run 8 processes on only one or another node, all is fine, but when i
run

mpiexec -n 16 -f machinefile siesta < input.fdf | tee log &

I get a lot of

Fatal error in PMPI_Comm_split: Other MPI error, error stack:
PMPI_Comm_split(418)..............: MPI_Comm_split(MPI_COMM_WORLD, color=2,
key=3, new_comm=0xd7e5a0) failed
PMPI_Comm_split(400)..............:
MPIR_Comm_split_impl(88)..........:
MPIR_Allgather_impl(744)..........:
MPIR_Allgather(705)...............:
MPIR_Allgather_intra(177).........:
MPIC_Sendrecv(186)................:
MPIC_Wait(528)....................:
MPIDI_CH3I_Progress(335)..........:
MPID_nem_mpich2_blocking_recv(891):
MPID_nem_tcp_connpoll(1830).......: Communication error with rank 1:


It is the same message with or without common filesystem shared by NFS

(by the way, NFS DRAMATICALLY slows down performance)


What did I wrong?

What is the right mpich2 device or process manager for Beowulf cluster
(several multiprocessor nodes, connected by eth cable)?



Mike Turchenkov
PhD Student, MEPHI, Moscow.

Responder a