Scheduling task graphs onto heterogeneous multiprocessors

In a heterogeneous system, the efficient exploitation of parallelism requires scheduling strategies that account for heterogeneity among processors to achieve an effective mapping of computations to processors. The paper presents a scheduling heuristic specialized for heterogeneous multiprocessor systems which make use of several different types of processors. The new algorithm is based on the greedy strategy: no processor remains idle if there is some task available that it could process. A graph called the classified typed task graph is used to describe the current status of tasks of different types at each scheduling step. With the graph, the algorithm dynamically evaluates the priorities of tasks only when there are multiply executable candidates, and then schedules an appropriate one onto the currently available processors of the corresponding type. A preliminary evaluation shows that this algorithm has promising performance