Hi Jeff,
> This is an old discussion. The "standard" job managers (meaning the
> ones that do not have 'lcg' in front of them) have the advantage that
> they have shared directories, supporting various forms of MPI that rely
> on shared homes.
>
> There is a well-publicized patch for the standard job manager (which has
> also been given to Globus, I don't know if they've taken it) -- this
> checks to see if the jobtype is multiple (meaning potentially MPI). If
> NOT, then it does a cd to $TMPDIR, so one gets the non-shared-home-dir
> functionality of lcgpbs when it's needed.
Indeed, and that patch should have been incorporated, but it still means
you have to mount the shared directories on every WN. In 2003 the most
urgent requirement brought up by LCG sites was to get rid of the need
for shared home directories, because the idea was that it would not scale
up to the hundreds or even thousands of WN some sites currently have.
The patch probably fixes the main issue, but there remains the risk of
all those WN grinding the NFS server to a halt under some conditions...
|