Multi-task scheduling framework for OpenCL programs on CPUs-GPUs heterogeneous platforms
Hao Henry Wang, Haofeng Wang, Sufang Wang · Third International Conference on Electronics and Communication; Network and Computer Technology (ECNCT 2021) · 2022
Heterogeneous systems consisting of multiple CPUs and GPUs are increasingly common as platforms for highperformance computing. OpenCL1 is widely used on this platform because of its cross-platform features and program portability. However, how to map OpenCL kernels onto the heterogeneous system in the presence of contention (i.e. multiple kernels compete for the computing resource) remains an outstanding problem. This is crucial to improve the efficiency of task execution. In this paper, we propose an efficient OpenCL task scheduling framework which schedules multiple kernels from multiple programs on CPUs-GPUs heterogeneous platforms. Our scheduling framework schedules kernels based on how well they match the actual running state of the current device. We show that the kernel execution is affected as the load increases. And we develop a novel model that schedule the kernel based on static and dynamic information about the kernel and the device. The framework provides adaptive and intelligent OpenCL multi-task scheduling on CPUs-GPUs heterogeneous platforms. We experimentally verified the efficiency of the framework. In the presence of resource competition, our approach achieves speedups of 1.47 and 1.61, compared to the two common scheduling strategies.