Scalability analysis of large codes using factorial designs

摘要

Analysis of scalability of parallel algorithm-architecture combination has been the subject of intense scrutiny for quite some time. This theoretical approach invariably requires detailed knowledge of the algorithm. Recently, Lyon and his coworkers at the National Institute of Science and Technology (NIST) developed an alternate black-box based approach to the analysis of scalability of large parallel codes. This approach is based on the time-honored principles from experimental design in statistics. Using this later approach, in this paper we analyze the scalability of a large code called Advance Regional Prediction System, which is a state-of the-art numerical weather prediction system on CRAY J-90 and IBM SP-2. This experimental approach does not require extensive knowledge of the underlying algorithm and can be automated.