Time Space Sharing Scheduling: A Simulation Analysis

We explain a new job scheduling class, called “Time Space Sharing Scheduling” (TSSS) for partitionable parallel machines. TSSS is a combination of time-sharing and space-sharing job scheduling techniques. Our proposed “Distributed Queue Tree” (DQT) is an instance of TSSS. We evaluate and analyze DQT behavior in more detail with a number of simulations. The result shows that DQT performs very well in low-load to high-load situations, almost independent of system size and task size distribution. We also compare our DQT and Scan Up batch scheduling, and we find that our DQT performs as well as Scan Up scheduling in processor utilization, but that both DQT and Scan Up have drawbacks in terms of scheduling fairness. Finally, we find that TSSS can inherently achieve higher processor utilization.