Upd:Linux Cluster: MPP Cluster and UV unavailable
publish at lrz.de
Di Feb 5 19:04:40 CET 2013
Changes to this message: MPP and UV unavailable
Dear users of the Linux Cluster systems at LRZ,
unfortunately both the MPP cluster and the Ultraviolet systems are
presently unavailable for user operation.
* On the UV, a defective NUMAlink component needs replacement.
* On the MPP cluster, we are observing failures in MPI message
transmission which need further investigation.
We'll keep you updated about the status of the above systems, and of
course are working toward a speedy return to user operation. Apologies
for the disruption of services.
Recent configuration changes:
* The default MPI environment on the MPP cluster will be changed from
Parastation MPI to Intel MPI. However, the mpi.parastation module
will remain available for legacy use until the end of 2013. On the
sgi ICE and UV systems, the sgi MPI (mpi.mpt) will remain default.
* The 8-way Myrinet Cluster will be retired from parallel processing,
and the nodes will be added to the serial processing pool. This
implies that the partition "myri_std" in the SLURM cell "myri" will
* For the serial queues, SLURM fair share scheduling will be
introduced. For the parallel queues, a combination of fair share
scheduling and favoring large jobs will be activated. This is to
prevent a single user from monopolizing cluster segments for long
times if there are many jobs in the queue.
* New storage systems will be introduced for the WORK (==PROJECT) and
SCRATCH file systems. Please note that LRZ will only migrate WORK
data to the new file system; data in SCRATCH will not be migrated.
However the old SCRATCH file system will remain available as a
separate mount in read-only mode on the login nodes until the end
of March, 2013. Migration of data can then be done via commands
cp -a my_scratch_subdirectory $SCRATCH
The environment variable $SCRATCH_LEGACY will remain defined and
point to the legacy scratch area until end of March, 2013.
This information is also available on our web server
Mehr Informationen über die Mailingliste aktuell