Actor-based parallel dataflow analysis

Abstract

Defining algorithms in a way which allows parallel execution is becoming increasingly important as multicore computers become ubiquitous. We present IFDS-A, a parallel algorithm for solving context-sensitive interprocedural finite distributive subset (IFDS) dataflow problems. IFDS-A defines these problems in terms of Actors, and dataflow dependencies as messages passed between these Actors. We implement the algorithm in Scala, and evaluate its performance against a comparable sequential algorithm. With eight cores, IFDS-A is 6.12 times as fast as with one core, and 3.35 times as fast as a baseline sequential algorithm. We also found that Scala's default Actors implementation is not optimal for this algorithm, and that a custom-built implementation outperforms it by a significant margin. We conclude that Actors are an effective way to parallelize this type of algorithm. © 2011 Springer-Verlag.