Hardware Reference
In-Depth Information
a multi-user, production system environment. Results presented here in doc-
umented daily performance samples from a Lustre file system on Stampede
as it transitioned from an empty file system at the beginning of production
to a fully utilized file system with periodic purging enforced and the natural
fragmentation that results.
As a fundamentally shared and interactive resource, the I/O subsystem
is often the primary avenue where users notice interrupts (whether they are
caused by the file system or not). It is also the source of the majority of system
interrupts by users who (often unknowingly) inject high-transaction workloads
over long periods of time that may impact performance or stability across the
resource. Consequently, TACC devotes significant effort toward monitoring
I/O usage at the application level. This monitoring process combines low-
level monitoring on each of the production I/O servers coupled with queries
into the underlying resource manager in order to delineate client I/O requests
on an individual job basis [2]. This allows TACC staff to identify applications
which may not be utilizing the I/O system effectively.
Finally, in order to improve application eciency and portability, TACC
staff continue to push community and individual research group applications
to adopt tuned I/O libraries that include parallel I/O semantics. In support
of this effort, TACC pro-actively promotes the use of libraries like HDF5 in
its scientific computing curriculum as a must-have component for researchers
to include in their application development toolbox.
Bibliography
[1] Peter J. Braam. The Lustre Storage Architecture. Technical report, Clus-
ter File Systems, Inc., 2003.
[2] John Hammond. Log Analysis and lltop. In Lustre User Group (LUG),
Orlando, FL, April 2011.
[3] Morris A. Jette, Andy B. Yoo, and Mark Grondona. SLURM: Simple
Linux Utility for Resource Management. In Lecture Notes in Computer
Science: Proceedings of Job Scheduling Strategies for Parallel Processing
(JSSPP) 2003, pages 44{60. Springer-Verlag, 2002.
[4] T. Minyard, K. Schulz, J. Boisseau, B. Barth, K. Milfeld, J. Cazes, J. Fos-
ter, S. Johnson, G. Jost, B.-D. Kimand L. Koesterke, and L. Wilson. Ex-
periences and Achievements in Deploying Ranger, the First NSF Path to
Petascale System. In Proceedings of the 2008 TeraGrid Conference, Las
Vegas, NV, June 2008.
