Efficient scheduling heuristics for GridRPC systems

In this paper, we examine several scheduling heuristics for GridRPC middleware relying on the time-shared model (a server can execute more than one task at a time). Our work is based on a forecast module called the 'historical trace manager' (HTM), which is able to predict durations of tasks in the system. We show that the predictions performed by the HTM are very accurate. The five proposed scheduling heuristics use these predictions to map submitted tasks to servers. Experimental simulation results show that they are able to outperform the well-known MCT heuristic for several metrics (makespan but also sumflow, max-stretch, etc.) and therefore provide a better quality of service for the client.