|Date||May 5, 2006|
|Speaker||Jeremy Kepner (MIT - Lincoln Laboratory)|
|Topic||The DARPA High Productivity Computing Systems (HPCS) Program and the Path to Usuable Petascale Computing|
High Performance Computing has seen extraordinary growth in peak performance from Megaflops to Teraflops in the past decades. This increase in performance has been accompanied by a large shift away from the original national security user base of the 1970s and 1980s to more commercially oriented applications (e.g., bioinformatics and entertainment). In addition, there has been a significant increase in the difficulty of using these systems, which is now the domain of highly specialized experts.
In response to these trends the DARPA High Productivity Computing Systems (HPCS) program was established to produce a new generation of economically viable high productivity computing systems for the national security and for the industrial user communities. The primary technical goals of the program are to produce petascale computers that can better run national security applications and are usable by a broader range of scientists and engineers. The HPCS program is fostering many technological innovations. One of the most important is the concept of a flatter memory hierarchy, which means that data from remote processors can be retrieved and used very efficiently. A flatter memory hierarchy will result in a significant performance increase (up to 2000x) in certain important classes of applications. In addition, a flatter memory hierarchy is much easier to program because the user doesn't have to worry as much about precisely tailoring their application to avoid the high cost of retrieving data.
To measure the memory hierarchy the HPCS program has developed the HPC Challenge benchmark suite and has sponsored the HPC Challenge contest that awards a prize for the best performance on each benchmark. In addition, there is also a coding contest, which rewards the best and most clearly written implementations of the benchmarks. This talk will discuss the results of the first HPC Challenge contest, as well as provide an overview of the HPCS program.
*This work is sponsored by the Defense Advanced Research Projects Administration under Air Force Contract FA8721-05-C-0002. Opinions, interpretations, conclusions, and recommendations are those of the author and are not necessarily endorsed by the United States Government.
We thank the generous support of MIT IS&T, CSAIL, and the Department of Mathematics for their support of this series.