Achieving Efficient Work-Stealing for Data-Parallel Collections

In modern programming high-level data-structures are an important foundation for most applications. With the rise of the multi-core era, there is a growing trend of supporting data-parallel collection operations in general purpose programming languages and platforms. To facilitate object-oriented reuse these operations are highly parametric, incurring abstraction performance penalties. Furthermore, data-parallel operations must scale when used in problems with irregular workloads. Work-stealing is a proven load-balancing technique when it comes to irregular workloads, but general purpose work-stealing also suffers from abstraction penalties. In this paper we present a generic design of a data-parallel collections framework based on work-stealing for shared-memory architectures. We show how abstraction penalties can be overcome through callsite specialization of data-parallel operations instances. Moreover, we show how to make work-stealing fine-grained and efficient when specialized for particular data-structures. We experimentally validate the performance of different data-structures and data-parallel operations, achieving up to 60X better performance with abstraction penalties eliminated and 3X higher speedups by specializing work-stealing compared to existing approaches.