Compiler and Run-Time Support for Improving Locality in Scientific Codes

Abstract

Modern microprocessors provide high performance by exploiting data locality with carefully designed multi-level caches. However, advanced scientific computations have features such as adaptive irregular memory accesses and large data sets that make utilizing caches effectively dificult. Traditional program transformations are frequently inapplicable or insuficient. Exploiting locality for these applications requires compile-time analyses and run-time systems to perform data layout and computation transformations. Run-time systems are needed because many programs are not analyzable statically, but compiler support is still crucial both for inserting interfaces to the run-time system and for directly applying program transformations where possible. Cooperation between the compiler and run-time will be critical for advanced scientific codes. We investigate software support for improving locality for advanced scientific applications on both sequential and parallel machines. We examine issues for both irregular adaptive and dense-matrix codes.