Experiences with implementing task pools in chapel and X10

Abstract

The Partitioned Global Address Space (PGAS) model is a promising approach to combine programmability and performance in an architecture-independent way. Well-known representatives of PGAS languages include Chapel and X10. Both languages incorporate object orientation, but fundamentally differ in their way of accessing remote memory as well as in synchronization constructs and other issues of language design. This paper reports on and compares experiences in using the languages. We concentrate on the interplay between object orientation and parallelism/distribution, and other issues of coding task parallelism. In particular, we discuss the realization of patterns such as objects that internally contain distributed arrays, and suggest improvements such as support for activity-local and place-local data, as well as scalar variable-based reduction. Our study is based on Unbalanced Tree Search (UTS), a well-known benchmark that uses task pools. © 2014 Springer-Verlag.