SuperMUC PetaFlop System: Status and schedule for start of user operation
LRZ aktuell
publish at lrz.de
Mo Jun 4 17:17:07 CEST 2012
Dear users of the national supercomputing facilities at LRZ,
With the installation of the new three Petaflop/s system "SuperMUC" by
IBM nearing its completion, we are now starting to prepare for user
operation of the system. The LINPACK benchmark has been successfully
run, and the results will soon be published at the International
Supercomputing Conference (ISC12) in Hamburg; people participating in
ISC are cordially invited to the Gauss Centre s booth, where further
information will be forthcoming. Also, LRZ s director Prof. Arndt Bode
will be speaker of a keynote talk at ISC12, an important topic of which
will be the innovations that were developed to enable energy efficient
usage of SuperMUC.
ISC12 event calendar
* Monday 9:00 ISC Opening session with 39th TOP500 list
* Thursday 12:00 keynote talk by Prof. Bode on extreme energy
efficiency.
* Gauss Centre for Supercomputing workshops
The Gauss Centre's booth at ISC12 has the number 150.
Starting user operation
It is presently expected that initial user operation with users
selected from large-scale projects will start at the end of June or the
beginning of July. These users will be contacted by LRZ approximately
one week in advance of being authorized for access. Some time later,
LRZ will proceed to ramp up usage until regular user operation will be
reached approximately in mid-August.
Limitations of early user operation
We expect that during early user operation we still need to flush out
glitches. Therefore, early users must expect unannounced or very
short-term scheduled maintenances. Early users are encouraged to report
any observed problems via the LRZ service desk.
Differences between SuperMIG and SuperMUC usage
* Node architecture: SuperMUC is comprised of 16-way Sandy Bridge
nodes; these are equipped with 32 GBytes of main memory much less
than the SuperMIG nodes. Due to OS requirements, probably 26 GBytes
per node will be available in user space. LRZ will also allow usage
of logical hardware threads (SMT/Hyperthreads, enabling use of up
to 32 tasks per node), however this resource is also partially used
by OS services, and it is very much application dependent whether
any performance improvement can be derived from use of SMT.
* Code compatibility: It is recommended to rebuild separate
executables for use on SuperMUC. While the compilers and basic
libraries will be the same as on SuperMIG, the new architecture
supports additional instructions (AVX SIMD extension), and IBM has
implemented a major new version of the parallel environment
(including the MPI library). Both of these may inhibit backward
compatibility, and we d like to avoid having to deal with a lot of
incidents reporting such (not really solvable) compatibility
issues.
* File systems: The file systems pointed to by the $SCRATCH and $WORK
environment variables will be the new, and vastly larger and
better-performing GPFS file systems. The HOME file system will be
still shared between SuperMIG and SuperMUC. We will provide
additional information on how to transfer data between the
non-shared SuperMIG and SuperMUC file systems via the file system
documentation.
Documentation
The LRZ SuperMUC web page will be updated to reflect the changes
discussed above. In particular the new LoadLeveler job classes and the
file systems will be documented; it may take some time before other
parts of the documentation are complete.
Courses
If you are interested in an introduction on how to use the features and
software of the new system, please register for the course
"Introduction to SuperMUC" scheduled for the week starting July 2.
Specialists from Intel and IBM have been invited to do presentations at
this event.
This information is also available on our web server
http://www.lrz-muenchen.de/services/compute/supermuc/aktuell/ali4325/
Reinhold Bader
Mehr Informationen über die Mailingliste aktuell