问题
I have a simple MPICH program in which processes send & receive messages from each other in a Ring order.
I've setup to 2 identical virtual machine, and made sure network is working fine. I've tested a simple MPICH program both machines and it works fine.
The problem arises when I try to communicate between processes on different machines like the above program. I'm getting the following error:
Fatal error in MPI_Send: A process has failed, error stack:
MPI_Send(171)...............: MPI_Send(buf=0xbfed8c08, count=1, MPI_INT, dest=1,
tag=1, MPI_COMM_WORLD) failed
MPID_nem_tcp_connpoll(1826): Communication error with rank 1: Connection refused
- SSH is passwordless & works fine on both sides.
/etc/hostsis configured properly.- Firewall is disabled on both machines.
- Configured NFS Client/Server and shared a directory between them. (According to this)
- Tried both MPICH & OpenMPI with Hydra
回答1:
Here what i did, And it works!
Installed following package using source code (tarball)
hydra
openmpi
Created hosts file (both node)
# cat /home/spatel/mpi/hydra/hosts
node1
node2
Set variable in .bashrc on (both node)
echo HYDRA_HOST_FILE=/home/spatel/mpi/hydra/hosts >> ~/.bashrc
Use HelloWorld MPI program to run on single node.
node1# /home/spatel/mpi/hydra/bin/mpiexec -np 1 /home/spatel/mpi/mpi_hello_world
Hello world from processor node1.example.com, rank 0 out of 1 processors
Run on multiple node using -machinefile option -np is number of processor
node1# /home/spatel/mpi/hydra/bin/mpiexec -np 4 -machinefile /home/spatel/mpi/hydra/hosts /home/spatel/mpi/mpi_hello_world
Hello world from processor node1.example.com, rank 0 out of 1 processors
Hello world from processor node2.example.com, rank 0 out of 1 processors
Hello world from processor node1.example.com, rank 0 out of 1 processors
Hello world from processor node2.example.com, rank 0 out of 1 processors
来源：https://stackoverflow.com/questions/14571203/mpich-communication-failed