On Aug 2, 2011, at 12:04 , Hans Lind wrote: > When doing large phonon calculations (dozens of atoms, distributed > over 1000s of nodes) > ph.x ends up doing a very large amount of disk i/o [...]. I have > tried setting "reduce_io='low'"
it works only for pw.x > Or do I have to resort to using distributed disk systems? It depends. If you have calculations involving "dozens" of atoms, you can perform a single irrep/wave-vector calculation on each node, working on local disks. Then, you have just to perform a final run that "collects" all the pieces. The latter involves only partial dynamical matrices, i.e. a small amount of data. The infrastructure to do this is in place in all recent versions, since it has been use for grid computing. > Are the developers working on some way to reduce the disk i/o in > future versions? there are projects of extending linear-response routines towards large systems. Interested in joining the effort? Note that there are only two ways to reduce I/O: - store things into memory - recalculate things instead of storing them In the first case, you need more RAM; in the second, more CPU. No way out. P. --- Paolo Giannozzi, Dept of Chemistry&Physics&Environment, Univ. Udine, via delle Scienze 208, 33100 Udine, Italy Phone +39-0432-558216, fax +39-0432-558222
