MRI: A PC-based Multithreading Environment for Solving Grand Computational Problems in Science and Engineering

EIA-9977508

Sohn, Andrew

McHugh, James A.

New Jersey Institute of Technology

MRI: A PC-based Multithreading Environment for Solving Grand Computation Problems in Science and Engineering

The purpose of this proposal is to explore an alternative to the expensive parallel computing, called cluster computing. The main idea is to develop and build a low-cost yet powerful parallel computer that consists of inexpensive dual-processor personal computers connected via commodity switchers and network cards. To overcome the long communication latencies associated with general purpose switches, it is proposed to combine the multithreading and communication co-processor capability. In this approach, each processor of a dual-processor PC has itself a dual role, acting both as a computation processor and a communication processor. Using software-enabled fine-grain multithreading, the duality of each processor allows the overlap of computation with communication, thus masking the long communication. The proposed architecture is designed to provide performance scalability and cost scalability. The principal result of the proposed research will be a cluster of PCs which can be used either as a powerful parallel computer to provide capabilities for computational science and engineering problems, or as a stable, scalable Web server as would be suitable for electronic commerce, and which at the same time can be used as a desktop computer for daily computing needs.