Dear all,
I am developing a system to manage MPI tasks on top of MPI. The
architecture is rather simple, I have a set of scheduler processes which
takes care to manage the resources of a node. The idea is to have 1 (or
more) of those scheduler allocated on each node of a cluster and then
create new MPI processes (on demand) as computation is needed.
Allocation of processes is done using MPI_Spawn.
The system now works fine on a single node by allocating the main
scheduler using the following mpi command:
mpirun --np 1 ./scheduler ...
Now when I scale to multiple nodes problems with default MPI behaviour
starts. For example lets assume I have 2 nodes with 8 cpu cores each. I
therefore set up a machine file in the following way:
s01 slots=1
s02 slots=1
and start the node schedulers in the following way:
mpirun --np 2 --machinefile machinefile ./scheduler ...
This allocates the processes correctly, now the problem starts when I
invoke MPI_Spawn. basically MPI spawn also uses the informations from
the machinefile and if 4 MPI processes are spawned 2 are allocated in
s01 and 2 on s02. What I want is to allocate the processes always in the
same node.
I tried to do this by specifying an MPI_Info object which is then passed
to the MPI_Spawn routine. I tried to set the "host" property to the
hostname of the machine where the scheduler is running but this didn't
help.
Unfortunately there is very little documentation on this.
Thanks for the help,
Simone
