Because of advances in commodity processors and networking technology, parallel computing platforms can now consist of cheap, commodity-designed symmetric multiprocessors (SMPs) as well as clusters of SMPs, workstations, and PCs. This research addresses methods to analyze, organize and schedule computing in such clusters which may be connected by a fast system-area network. Research activities being carried out include: (1) Designing and developing analytical and experimental performance models and tools to quantitatively guide application users and computer architects to optimize design decisions for a given budget, and a given class of application programs. Focus is on constructing a new cluster, upgrading an existing cluster, and designing a memory hierarchy of a cluster. 2) Developing and implementing both static and dynamic scheduling schemes, as well as aiming at coordinating parallel tasks for load balancing and effectively using heterogeneous workstation resources. (3) Integrating the performance prediction models and scheduling schemes into a visualization environment.
