Section: New Results

Heterogeneous task scheduling

Taking advantage of heterogeneous systems requires to carefully choose which tasks should be accelerated. Simple heuristics allow to get fairly good performance, but do not have approximation ratio that would provide performance guarantees. The ROMA Inria team designed advanced heuristics which do have approximation ratios. We have implemented one within StarPU and indeed improved the performance over existing heuristics. The implementation required to revamp part of the StarPU toolkit dedicated to writing scheduling heuristics.