Have Your Cake and Eat It? Productive Parallel Programming via Chapel’s High-level Constructs

Abstract

Explicit parallel programming is required to utilize the growing parallelism in computer hardware. However, current mainstream parallel notations, such as OpenMP and MPI, lack in programmability. Chapel tries to tackle this problem by providing high-level constructs. However, the performance implication of such constructs is not clear, and needs to be evaluated. The key contributions of this work are: 1. An evaluation of data parallelism and global-view programming in Chapel through the reduce and transpose benchmarks. 2. Identification of bugs in Chapel runtime code with proposed fixes. 3. A benchmarking framework that aids in conducting systematic and rigorous performance evaluation. Through examples, I show that data parallelism and global-view programming lead to clean and succinct code in Chapel. In the reduce benchmark, I found that data parallelism makes Chapel outperform the baseline. However, in the transpose benchmark, I found that global-view programming causes performance degradation in Chapel due to frequent implicit communication. I argue that this is not an inherent problem with Chapel, and can be solved by compiler optimizations. The results suggest that it is possible to use high-level abstraction in parallel languages to improve the productivity of programmers, while still delivering competitive performance. Furthermore, the benchmarking framework I developed can aid the wider research community in performance evaluations.