| Abstrakt: |
Optimally mapping diverse independent tasks onto machines in a distributed system, in general, is an NP hard problem. Further, distributed systems augmented with GPUs are quite different from those without GPUs. In order to better utilize machines with both CPUs and GPUs, dynamic load balancing has been introduced and it can significantly affect the scheduling performance of such distributed systems. In this paper, we proposed four heuristics, adapted them into a set of common assumptions, added dynamic load balancing, and compared their testing results based on the simulation experiments. The performance of these heuristics, as well as the effect of task redistribution, was carefully studied. MCT with task redistribution out-performs other algorithms. Task redistribution can significantly improve performance, especially when the speed of CPU and GPU is close to each other. [ABSTRACT FROM AUTHOR] |
