Load and Resource Management
(Special Focus: Parallel I/O)
November 29-30, 2004 Cluster systems form an important technical basis of high performance computing. However, their efficient usage is still not at all easy. We see various problems with system software, with libraries, and with application software. In the last decade many of them were tackled successfully such that we can get a high compute performance out of these systems. This is reflected by the fact that more than half of the systems shown in the Top500 list are clusters.
An issue that is still under heavy investigation is input/output (I/O). We see parallel file systems that offer high I/O performance at the system level and parallel file I/O access routines like those defined in the I/O section of the MPI2 specification. Achieving scalabale optimal performance in a cluster system with many disks is still a critical issue. Various techniques and optimizations are applied. Despite these efforts we do not see too many parallel I/O systems and not too many programs already being adapted to parallel I/O.
The goal of the workshop is to bring together researchers in the field of parallel I/O and to discuss current trends and ideas for future work and possible cooperations between the participants of the workshop.
The workshop series started in March 1997 at the Medizinische Universität zu Lübeck. Information and abstracts of talks at previous meetings can be found at silr.de.
This year we will have Dr. Rob Ross from the Argonne National Laboratories as invited guest. Dr. Ross will give a talk on the current research in the parallel file system PVFS2.
Topics that should be discussed in more detail during the workshop include but are not limited to:
If you like to participate actively in the workshop please send a half page abstract of your planned presentation to email@example.com before November 15. We will reply with a notification of acceptance by November 18.
- Parallel file system implementation
- Parallel file system deployment
- Efficient meta-data handling
- Load balancing in parallel file systems
- Visualization of load characteristics with parallel I/O
- Network support for parallel I/O
- I/O clusters vs. compute clusters
- Fault-tolerance in parallel file systems
We do not plan to produce any proceedings or publications. Instead copies of the slides will be made accessible to all participants. There is an implicite non-disclosure agreement: we would like people to speak about there current research and results and we will not make the slides available to the public.
The registration fee for the workshop is 80 Euro.
Im Neuenheimer Feld 348
Room: to be announced
Maps are here