Task-to-Accelerator Mapping for Heterogeneous Systems Using Heuristics

Heterogeneous hardware platforms can be used to improve the performance of computing systems in terms of application response time. They are in high demand, especially with the recent emergence of complex AI applications. Parallel applications can be used on these platforms to utilize most of the hardware system capacity and decrease the execution time. However, predictability is still a notable challenge on these platforms when used in time-critical systems due to the variability in the execution of parallel runtime systems, like OpenMP. Therefore, this paper proposes a new heterogeneous task-to-accelerator mapping approach using heuristics to improve the performance and predictability of OpenMP applications running on heterogeneous systems.