Hi,
I 'm doing some tests in the Memento cluster of the University of Zaragoza
on TiC system with a k- 100k pts , 4 nodes with 64 CPUs per node. It is a
system that does not share RAM and hard disks between nodes during
calculations. Initially the parallel computation with Wien2k stopped in the
Hello César,
To perform parallel calculations you do need a shared directory between
all nodes. As you have described '/home' appears to be a form of shared
storage.
What its intention is, is of course not well-known to us. If it is
shared there is no direct reason it cannot function for
It is getting complicated when you do both MPI + k-point
parallelization. In large calculations there is usually less k-points. Will
it be possible to test MPI with the local scratch without k-point
parallelization (i.e., k-point run sequentially)? This will help to mediate
problems mentioned by
3 matches
Mail list logo