Advanced Static and Dynamic Scheduling of HPC Applications on Petascale Computer Systems with GPU Accelerators

The DOE undertakes scientific research in areas that are extremely computation intensive. These areas include climate modeling, nuclear physics, high energy physics, biomedical engineering, and combustion chemistry. DOE¿s supercomputer systems provide performance in the 100¿s of TeraFLOPS (TFLOPS), with plans to achieve PetaFLOPS computing levels by 2009. One particularly cost effective approach to boosting performance/power is the use of ¿accelerators.¿ The most prominent example is inexpensive graphics processors that provide huge performance advantages in terms of GFLOPS and memory bandwidth. However, this performance comes at a cost: these systems are notoriously difficult to program and the resulting programs lack portability. One of the key bottlenecks is the development of effective software programming tools and runtime libraries for Graphics Processing Units (GPUs). This project will develop compiler technologies to facilitate the programmability of these processors. In particular, a class of algorithms, each incorporating multiple GPUs, will be mapped automatically to clusters of computers. The key technical challenge involves extracting several degrees of parallelism and locality, and generating correct communications and synchronizations, which now must be done manually. Commercial Applications and other Benefits as described by the awardee: The technology should benefit the DOE national labs with their demanding needs for high performance computing. Another potential client is DARPA¿s High Productivity Computing Systems (HPCS) initiative, an interagency program that supports the development of PFLOPS-class computer systems that are substantially easier to program and use than current systems.