Jeff Squyres ha scritto:
> LAM uses a package called ROMIO for its MPI-2 I/O functionality.
>
> Rajeev, one of the ROMIO maintainers, is a member of this list.
> Rajeev -- can you comment on what is happening here?
>
>
> On Jan 23, 2006, at 3:42 PM, Barry A. Croker wrote:
>
>
>>Hello
>>I've been working on using parallel I/O with FORTRAN 90. My code is
>>working fine on my workstation, but I'm getting weird errors on our
>>linux cluster writing to an NFS mounted volume. The program writes
>>restart files using parallel I/O. The restart files are generally
>>about 13 Mb each. Before I close the file I use MPI_FILE_GET_SIZE
>>and the size is reported correctly for all processes which are
>>writing the file (~13 Mb). However, when I list or copy or move the
>>file on the NFS volume is is reported as 641 Gb (!). Strangely
>>enough, when I do a df on the filesystem the massive file does not
>>seem to appear, as the total disk usage is reported around 300 Gb.
Just a comment from an OS point of view: it is possible that the filesystem you
use packs the blocks which contain only zeroes, so that your file is actually
such big but full of zeroes and it does not occupy much physical space on disk.
Davide
|
