Dynamic Scheduling of Independent Tasks with a Budget Constraint on a Desktop Grid

In an Internet desktop grid, the computing power may be leveraged by providing economic incentive for computer owners. Using batch mode scheduling heuristics, such as MinMin, MaxMin or Sufferage, this paper describes an algorithm that performs independent task scheduling on a large scale heterogeneous system, with a budget constraint. Under the restriction of never exceeding a periodically refreshed budget, successive optimizations are performed to enhance the makespan by finding a suitable distribution of the budget among the given tasks, through the given time horizon. For the scalability purpose, the computing resources are partitionned according to their speed into classes. First, a schedule on these classes is computed using one of the mentioned well-known heuristics, and then, tasks assigned to a class are actually scheduled to the respective computing resources, as these tasks were ranked by the first scheduling algorithm. To evaluate the effectiveness of our scheduling approach, we have built a simulator that uses real traces obtained from the BOINC based XtremLab project. The results show that the repartition of the budget among tasks, and through time, improves the makespan by a significant factor. They also show that the adaptation to the grid scale can achieve comparable makespan when the availability interval of computing node is not too small.