Page 1 of 1

Serpent on a cluster

Posted: Wed Mar 25, 2015 4:54 pm
by Domenico
Hi to all,
I am trying to run Serpent on a cluster, but I have faced with several problems about RAM consume. Talking with the administrator of the machine about a possibile solution with a large numbe of tasks MPI, could be helpful for us knowing how Serpent manages the ram on the threads(master and slaves). Can you tell me how Serpent allocates the memory using a MPI parallelization? Because we have checked that the entire job requests less than memory on a single node, but the code is stopped in the fase of sharing the MPI tasks

Thanks you,

Domenico

Re: Serpent on a cluster

Posted: Thu Mar 26, 2015 12:08 pm
by Jaakko Leppänen
The problem with MPI parallelization is that each task requires its own copy of the geometry, cross section data, etc., so if you are running the calculation with multiple tasks within the same node, the memory demand is correspondingly multiplied.

This is the main reason we started working on Serpent 2 with hybrid OpenMP + MPI parallelization, which allows distributing the calculation between nodes with MPI and using shared-memory OpenMP inside each node. So the solution here is to move to Serpent 2. Send me an e-mail and I'll send you the software license agreement for Serpent 2.

Re: Serpent on a cluster

Posted: Thu Aug 20, 2015 3:49 pm
by Giorgio Baiocco
Hello,
I am trying to run Serpent 1.1.19 on a cluster but I get some problems. In particular I would like to run the calculation on two different nodes (4 slots each one) in parallel. I type:

mpirun -np 8 --hostfile host sss assembly.inp

where host is the file where I specify the nodes.

The calculations starts and 8 processes are created (4 per node). The problem occurs at:

Sharing input data to MPI tasks...

After that the calculation doesn't go on anymore. I don't understand if the issue is with Serpent or with some set on MPI.

Can anyone help me?


Thank you,

Giorgio