With a data parallel design, GPUs depend on uniform work distribution to expose their full potential. Therefore, irregular applications suffer from serious performance degradation as it is highly challenging to schedule irregular tasks on a GPU: It requires understandings of GPU architecture and irregular applications to devise a scheduling most suitable in this context, not to mention error-prone concurrent programming. This paper proposes a two-level scheduling to distribute irregular tasks and enable resource sharing on GPUs, by managing tasks and threads hierarchically. Meanwhile, we manage to group cache friendly tasks for more data reuse in L1 cache. We further extend our scheduling to handle nested irregularities. Besides, we devise a programming framework to facilitate the task scheduling for application programmers. The experimental results show that our approach effectively improves performance of six irregular applications on a typical platform, yielding a harmonic-mean speedup of $$2.1\times $$2.1× at a small schedule cost, and does not burden programmers with lots of work.
Read full abstract7-days of FREE Audio papers, translation & more with Prime
7-days of FREE Prime access