merged wannier center and coordinate output file
akoh... at gmail.com
Mon Jul 9 18:11:08 UTC 2007
On Jul 9, 1:58 pm, Teodoro Laino <teodor... at gmail.com> wrote:
> just to share my experience:
> on our (CSCS) XT3 machine IT people installed a special library
> (iobuf, vendor libraries from CRAY)..
> this library does what you were describing in your e-mail.. is
i was referring exactly to this module/library.
> intercepting all IO operations and keeps them in memory and only when
> the IO buffer is full (~ 100 MB per node) everything is flushed.. in
> this way all the problems related to the scalability of the code due
> to IO operations are highly dumped..
right! it is very useful. i've been using even larger buffers
in some cases. but open/close interferes with it since iobuf
cannot know which close is the last one, so one can only
intercept closing of files that one does not care about
after the run is done. so for all files that one wants to
keep, iobuf is of limited efficiency, since every close implies
a flush and a sync.
as i wrote before, it is of limited relevance now, but i expect
this to become more of a problem in the not so far future.
> On 9 Jul 2007, at 19:31, Axel wrote:
> > this may not be a big issue for most of the current machine and
> > users, but i expect that more and more machines will have to use
> > parallel file systems like lustre, GPFS and alike and at the same
> > time, people want to run larger jobs faster on their new fancy
> > machines,
> > so this may become a much more important over the next few years.
More information about the CP2K-user