A hybrid static/dynamic approach to scheduling stream programs

Abstract

Streaming languages such as Streamlt are often utilized to write stream programs that execute on multicore processors. Stream programs consist of actors that operate on streams of data. To execute on multiple cores, actors are scheduled for parallel execution while satisfying data dependencies between actors. In StreamIt, the compiler analyzes data dependencies between actors at compile-time and generates a static schedule that determines where and when actors are executed on the available cores. Statically scheduling actors onto cores results in no scheduling overhead at runtime and allows for sophisticated compile-time scheduling optimizations. Unfortunately, static scheduling has a number of severe limitations. The generated static schedule is inflexible and cannot be adapted to run-time conditions, such as cores that are unexpectedly unavailable. Static scheduling may also incorrectly load-balance cores due to inaccurate static work estimates. This thesis contributes a hybrid static/dynamic scheduling approach that attempts to address the limitations of static scheduling. Dynamic load-balancing is utilized to adjust the static schedule to run-time conditions and to correct load imbalances that might exist after static scheduling. Dynamic load-balancing is designed to add very little run-time overhead.