Understanding Parallelization Tradeoffs for Linear Pipelines

Abstract

Pipelining techniques execute some loops with cross-iteration dependences in parallel, by partitioning the loop body into a sequence of stages such that the data dependences are not violated. Obtaining good performance for all kinds of loops is challenging and current techniques, e.g., PS-DSWP and LBPP, have difficulties handling load-imbalanced loops. Particularly, for loop iterations that differ substantially in execution time, these techniques achieve load-balancing by assigning work to threads using round-robin scheduling. Algorithms that rely on work-stealing e.g., Piper, efficiently handle load-imbalanced loops, but the high overhead of the scheduler implies poor performance for fine-grained loops. In this paper, we present Proteas, a programming model to allow tradeoffs between load-balancing, partitioning, mapping, synchronization, chunking, and scheduling. Proteas provides a set of simple directives to express the different mappings to handle a loop's parallelism. Then, a source-to-source compiler generates parallel code to support experimentation with Proteas. The directives allow us to investigate various tradeoffs and achieve good performance according to PS-DSWP and LBPP. In addition, the directives make a meaningful comparison to Piper possible. We present a performance evaluation on a 32-core system for a set of popular pipelined programs selected from three widely-used benchmark suites. The results show the tradeoffs of the different techniques and their parameters. Moreover, the results show that efficient handling of load-imbalanced fine-grained loops remains the main challenge.