System Support for Implicitly Parallel Programming

Implicit parallelization involves developing parallel algorithms and applications in environments that provide sequential semantics, e.g., the C programming language. System tools convert the parallel algorithms into a set of threads partitioned appropriately for a particular parallel machine organization. The resulting parallel programs are easier and faster to develop, debug and maintain, because the programmer can request a meaningful and well defined program state at any point of execution. The contribution of this paper is a case study of a video encoding application. We show that error checking code, code reuse, and variable scoping interfere with parallelization. We suggest that system tools must perform reactive and speculative transformations if they are to reduce this tension between application robustness and parallelization.