Abstract
The paper proposes compilation techniques for block-cyclic data redistribution on distributed-memory multicomputers when two-level data-processor mapping is adopted. Since static data distribution is in-suucient in achieving high performance computing for many scientiic and engineering applications, dynamic data redistribution is the only alternative to get better performance. However, data redistribution is costly at runtime. How to reduce the indexing overhead and communication overhead is very crucial for distributed-memory multicomputers. This paper proposes compilation techniques to eeciently generate communication sets for block-cyclic data redistribution , assuming data-processor mapping is a two-level mapping, that is, related array objects are rst aligned with each other or aligned with an abstract index space; this group of arrays is then distributed onto the user-declared abstract processors. The proposed method can save memory space by compressing holes caused by two-level data-processor mappings and eliminate redundant communication.