Methods for multitasking among real-time embedded compute tasks running on the GPU

摘要

In this study, we provide an extensive survey on wide spectrum of scheduling methods for multitaskingrnamong graphics processing unit (GPU) computing tasks.We then design several schedulersrnand explain in detail the selected methods we have developed to implement our schedulingrnstrategies. Next, we compare the performance of schedulers on various workloads running onrnFermi and Kepler architectures and arrive at the following major conclusions: (1) Small kernelsrnbenefit from running kernels concurrently. (2) The combination of small kernels, high-priorityrnkernels with longer runtimes, and lower-priority kernels with shorter runtimes benefits from arnCPU scheduler that dynamically changes kernel order on the Fermi architecture. (3) Because ofrnlimitations of existingGPU architectures, currently CPU schedulers outperform their GPU counterparts.rnWe also provide results and observations obtained from implementing and evaluatingrnour schedulersontheNVIDIAJetsonTX1system-on-chip architecture.Weobserve that althoughrnTX1 has the newer Maxwell architecture, the mechanism used for scheduler timings behaves differentlyrnon TX1 compared to Kepler leading to incorrect timings. In this paper, we describe ourrnmethods that allow us to report correct timings for CPU schedulers running on TX1. Finally, wernpropose new research directions involving the investigation of additional scheduling strategies.