How Moderate-Sized Risc-Based SMPs Can Outperform Much Larger Distributed Memory MPPs

摘要

Traditionally, comparisons between computer systems have been based primarily on theoretical peak performance. While many people still take this approach, an increasingly common practice is to compare computer systems based on delivered levels of performance. This of course raises a whole host of questions about how one should go about measuring or estimating the delivered level of performance. However, even this approach has a fundamental problem. It assumes that all PLOPS are of equal value. When computational scientists were primarily using mainframes and vector supercomputers, this was probably a reasonable assumption. The same holds true when comparing the performance of large distributed memory MIMD MPPs. However, when comparing the algorithms of chocie used on these two classes of platforms, one frequently finds a singnificatn differnece in the number of FLOPS required to obtain a solution with the desired level of precision. Historically, this dichotomy was considered to be troubling, but largely unavoidable if one was to take advantage of the benefits offered by aprallel computing. Recently, however, a new class of computers has come onto the market. Moderate-sized RISC-based SMPs offer much of the scalability of modern large MPPs, while allowing one to parallelize efficient vecotr algorithms that were previously considered to be nonparallelizable. The net resutl is that for some problems a 128 processor Origin 2000 can deliver levels of performance that might require the use of a 500+ processor MPP using more traditional approaches.