[Pw_forum] EXX files I/O

Tyler Grassman tgrassma at ucsd.edu
Sun Aug 19 03:28:38 CEST 2007


Hello.  I've been working with the EXX implementation in PWscf for a little
bit now (just doing testing on bulk Si for the time being, getting to know
the code a little bit, as I'm new to PWscf).  I'm working on figuring out
the optimal parallelization and disk I/O parameters.  My cluster uses an NSF
file system to share the /home directory (I know, everyone says not to with
PWscf, but I don't see that I've much other choice).  However, the compute
nodes have large scratch disks, so setting the wfcdir tag to use the local
node scratch disks helps reduce the NFS disk I/O.  However, I find that the
prefix.exx# files are still being written to the outdir, and they tend to be
pretty big (each file is tens to hundreds of MBs, depending on the number of
procs the job is being run on, even for this small bulk system).  It looks
like these files get rewritten every hybrid refinement loop (i.e. "NOW GO
BACK TO REFINE HYBRID CALCULATION").  I'm guessing these files will get
considerably larger once the system gets larger, too, in which case this I/O
will really end up costing a lot in the wall time (and I plan on trying some
considerably larger systems in the near future), yes?  I looked in input.f90
source file and didn't see anything about it, but I was wondering if there
was any way (or a plan to implement such a thing, if possible) to treat
these files in the same way (some equivalent to wfcdir)?  Or if not, does
anyone have any suggestions?

Thanks,

Tyler Grassman
Materials Science and Engineering
University of California, San Diego




More information about the users mailing list