Supercharging the APGAS Programming Model with Relocatable Distributed Collections

TL;DR

This paper introduces a relocatable distributed collections library built on APGAS for Java, enabling dynamic load-balancing and high-level control of distributed computations, with demonstrated performance benefits and programmability improvements.

Contribution

We developed a relocatable distributed collections library that extends APGAS for Java, supporting dynamic load-balancing and flexible data distribution through high-level abstractions.

Findings

Superior programmability demonstrated

Better performance on one benchmark

Reasonable overhead on second benchmark

Abstract

In this article we present our relocatable distributed collections library. Building on top of the AGPAS for Java library, we provide a number of useful intra-node parallel patterns as well as the features necessary to support the distributed nature of the computation through clearly identified methods. In particular, the transfer of distributed collections' entries between processes is supported via an integrated relocation system. This enables dynamic load-balancing capabilities, making it possible for programs to adapt to uneven or evolving cluster performance. The system we developed makes it possible to dynamically control the distribution and the data-flow of distributed programs through high-level abstractions. Programmers using our library can therefore write complex distributed programs combining computation and communication phases through a consistent API. We evaluate the performance of our library against two programs taken from well-known Java benchmark suites, demonstrating superior programmability, and obtaining better performance on one benchmark and reasonable overhead on the second. Finally, we demonstrate the ease and benefits of load-balancing and on a more complex application which uses the various features of our library extensively.