[Dock-fans] Question re MPI Dock
p.crawford at unsw.edu.au
Wed May 17 17:02:56 PDT 2006
Hi. Having a few problems with MPI Dock on large cluster.
The administrators request that jobs be run from a local folder, set up by
PBS, on the local file system of the Master node. This is to avoid slowdowns
on the head node. As far as I can tell, this local scratch space is not
accessible via NFS from the other execute nodes.
It seems to me that Dock 5 MPI version actually has each node read the
dock.in file and produce its own separate dock.out.nodeno file, though the
manual says that "The Master node performs file processing and input/output,
whereas the slaves perform the actual calculations".
In my case, when I run from this local scratch space, the execute directory
is not accessible from the other execute nodes and only the local processes,
in my case dock.out and dock.out.16 had been written. Though I had been
allocated 15 other DP nodes, Dock was not running on those I logged onto.
I'd like to confirm that I need to run the MPI job from a working directory
which is accessible from the other execute nodes. Am I correct?
Thanks in advance,
Philip Crawford, B. Comp. Sc., MIEEE
School of Medical Sciences
The University of NSW
Phone: +61-2-9385 2564
Mobile: +61-419-294 698
Fax: +61-2-9385 1059
Email: p.crawford at unsw.edu.au
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Dock-fans