JUGGLER: A DYNAMICWARP SCHEDULER FOR GPGPUS

Abstract

In the age of data-centric computation, the General-Purpose Graphics Processing Unit (GPGPU) has emerged as a vital tool for accelerating a wide range of applications, from graphics rendering and scientific simulations to the cutting-edge field of machine learning. The backbone of GPGPUs is their massively parallel processing and fast-switching threading architecture. The warp schedulers are the heart of such a computational model. In this work, our analysis shows that different GPGPU applications perform better with different warp scheduling modes, which intrigued us and led us to develop a runtime switching mechanism that can predict the best mode of execution at each execution stage of the running application. We ended up designing our Juggler that achieves an average performance gain of 0.83% over GTO, 8.42% and 13.34% over LRR and TL, respectively, on a Fermi architecture and 0.69%, 3.93%, and 3.12% improvement over GTO, LRR and TL using Turing architecture.