On 26/11/2014 09:13, Martin Sandve Alnæs wrote:
Chris: when jit is called on a subset of processors, the others
obviously cannot participate. The check for rank 0 within the group is
exactly to stop all the other participating processors from writing to
disk simultaneously. So if the instant cache is in a shared directory,
there's no more disk usage here.
Of course if you have two groups doing exactly the same operations,
then rank 0 from each group will race for copying into the instant
cache. In that case we don't have mpi communication between the racing
processors because they are on different groups, and we'll need the
file locking in instant to work. Which I guess is still a problem with
NFS...
Actually, this could work quite well for some HPC systems, which have
/tmp mounted locally (e.g. in memory) on each node.
In that case, if the instant cache is set up in /tmp, the nodes don't
touch the filesystem at all.
All we would need to do, is to be able to construct MPI groups based on
which processes share a compute node.
Chris
_______________________________________________
fenics mailing list
[email protected]
http://fenicsproject.org/mailman/listinfo/fenics
