Section: New Results
Work stealing scheduling algorithm taking care of communication
On some applications, the amount of data transfers can be high. To minimize the amount of data transfers during the execution, Jean-Noel Quintin has developed an algorithm called WSCOM which uses the DAG structure of the application. For each steal request, the work-stealing algorithm tries to balance the load between the thief and the stolen processor. Thus, WSCOM tries to divide the work on the stolen processor into two parts with a small number of edges between the two parts. This cutting is done with a negligible overhead at each steal request. This algorithm has been implemented in a tool called DSMake. This tool executes the set of tasks described by a Makefile on a distributed platform. In addition, I have developed a simulator to validate algorithm performance and its behavior. We compared WSCOM and severals static list-scheduling algorithms. The comparison shows that WSCOM outperforms list-scheduling algorithms, on clusters with some network congestion.
Besides, based on SIPS analysis of work stealing, Stefano Mor in his thesis compared the influence of the choice of the stolen tasks on the number of steal operations, distinguishing unsuccessful and successful steals. While standard bounds are related to unsuccessful steals, they are pessimistic with respect to the number of successful steals that define intensive data communications.