![[ IBM Research ]](http://www.research.ibm.com/research/newmast/projectmasts/masthead17.GIF)
IBM's HPC Challenge Class 2 Entry
The First HPC Challenge Class 2
Competition was held this year at the SC|05. The award was shared
by the IBM team and Cray.
The IBM entry consisted of a UPC
implementation of two benchmarks, Random
Access and Stream
Triad. The benchmarks were compiled using an experimental version of the IBM XLUPC
compiler and run on the BlueGene/L
machine.
In addition to the fact that this is the first time a UPC program
has been scaled to more than a few hundred threads, the performance
obtained using these simple codes is surpassed only by hand optimized
algorithms for these benchmarks.
The submission:
Summary of results:
| Threads | Random Access (GUPS) | STREAM (GB/s) |
| 1 | 0.00054 | 0.73 |
| 2 | 0.00078 | 1.46 |
| 64 | 0.02000 | 46.72 |
| 2048 | 0.56000 | 1472.00 |
| 65536 | 11.54000 | 47830.00 |
| 131072 | 16.72500 | 95660.00 |
Acknowledgements: This work was supported in part by DARPA Contract NBCH30390004. Our thanks go to the many other people who supported this project, In particular,
we would like to thank Roch Archambault, Anthony Bolmarcich, Jose Castanos,
Sid Chatterjee, John Gunnels, Manish Gupta, Roland Koo, Larry Lindsay, Fred Mintzer and Tom Spelce (LLNL).