Optimal scheduling policies for mixed scalar-vector multiprocessor supercomputers

A queuing network model is developed for multiprocessing supercomputer systems implemented on two heterogeneous processors, a scaler unit and a vector unit, which work concurrently on a single process. The model is analyzed theoretically for a representative case of model job mixes. It is found that the utilization of one of the two central service stations in the model can only be maximized at the expense of minimizing the utilization of the other central service station. The method of analysis adopted is the technique of variable aggregation based on nearly complete decomposability