Processor Grain Size and Overhead for Massive Parallelism
D. Scott Wills · 2005
*L? &:TA4 Abstract !l%w+pw presents a model of parallel system performance, ancludang the effect acal sequentaal sequences and parallel overhead. o surveys two archztectures whzch attempt to aze parallel overhead for communacatzon, synchronazataon, namang, and task and storage management. 1 Pursuit of Parallelism Parallel architectures have been studied for over twenty five years. The goal is harness the power of thousands of processing elements to obtain performance greater than can be obtained via a sequential system. A range of obstacles in algorithms, languages, compilation, networks, and processor design have thwarted this goal. This paper addresses one aspect of a parallel system: the design of the processing element. It defines a model to examine issues in parallel execution, including the effect of parallel overhead. It then surveys two architectures designed to minimize parallel overhead: the MIT MDP [a] and the Georgia Tech Pica [lo] architectures.