Section: New Results

Task Size Control with XcalableMP/StarPU

On the work sharing among GPUs and CPU cores on GPU equipped clusters, it is a critical issue to select the task computational weights suited to these heterogeneous computing resources. We have been developing a solution for this problem, based on the cooperation of a PGAS language named XcalableMP (developed at the University of Tsukuba) together with a runtime sytem named XMP-dev/StarPU building on the work of the University of Tsukuba and on the StarPU platform developed by the Inria Runtime Team. Through the development, we found the necessity of adaptive task weight control for the GPU/CPU work sharing to achieve the best performance for various application codes. In particular, the language was extended to add a new feature allowing to alter the task size to be assigned to these heterogeneous resources dynamically during application execution. As a result of performance evaluation on several benchmarks, we confirmed the proposed feature correctly works and perform well even for relatively small size of problems.