SC2000 Technical programme - drifting away from pure HPCN

Utrecht 30 November 2000 Traditionally, the technical program is mostly of high quality and this year this was at least true for the talks the author has attended. For the regular papers there were 2 parallel tracks of half hour talks that ran along with the Masterworks invited papers. As the field of HPCN has broaded over the years this was reflected in the session subjects.

Session subjects:

  1. Applications (2 sessions).
  2. Biomedical Applications.
  3. Visualisation.
  4. Compiler optimisation.
  5. Scheduling.
  6. MPI
  7. MPI/OpenMP.
  8. Cluster Infrastructure.
  9. QoS/Fault Tolerance.
  10. Networking.
  11. Hardware-based tools.
  12. Software tools.
  13. Data Grid.
  14. Application support.
  15. Grid middleware.
  16. Numerical algorithms.
  17. Parallel Programming
  18. Potpourri (3 talks that did not fit the other categories).
Comparing this list with those of earlier years one can already get an inkling of the shift in focus: HPF has disappeared from the agenda. In fact, it was only mentioned in the discussion during the talk about an extension of OpenMP to include Data Distribution directives. The extensions were made by Jonathan Harris et. al. to work on a Compaq AlphaServer SC system and did indeed improve the C-codes that were investigated but, as was rightly remarked from the audience, in this way OpenMP is given a distinct HPF flavour which is going against the grain of the original OpenMP ideas where one is postulating a shared memory and takes possible non-uniformity in data access for granted. Now the first experiences with actual codes appear drawbacks associated with TLB misses, etc. become apparent and ways are explored to evade them.

In another paper in the same session the influence of data placement by three memory page placement schemes were investigated: round robin, random, and worst case placement, i.e., all pages resided initially on one node and had to migrate according to the program's needs. The first two strategies turned out to be not very different but, not surprisingly, better than the worst case scheme by some tens of percents. The round robin and random schemes usually did somewhat better than the standard first touch approach as used by SGI (all experiments were conducted on an SGI Origin2000). However, the results were obtained for codes from the NAS Parallel Benchmarks and inherently relied on the iterative nature of the algorithms in these programs. This enabled the move_on_next_touch procedure for both strategies to be effective. This will of coarse not work for programs that lack the iterative structure of the considered programs.

Aad van der Steen

[News on Advanced IT][Calendar][Analysis][IT in Medicine]