Proportional Share Scheduling employing Performance-aware Virtual Time in Multiprocessor Systems

Abstract

In proportional share scheduling, the different performances of CPUs can make running tasks unfair during a period. To make them fair in a system, we present a proportional share scheduling employing performance-aware virtual time (PVT) maintained globally. This PVT is the share of CPU time received by a task and increases at a rate proportional to the performance of CPU where the task is running on and inversely proportional to the weight of the task. The schedulers of CPUs, when they assign CPU time to a task, utilize PVT to make a decision which task and how long it should preempt CPU to minimize the difference of PVTs among tasks. We evaluated our approach experimentally on general purpose operating system in the homogeneous and heterogeneous multiprocessor (HMP) systems. On both systems, the results show the significant improvement that is near-perfect (around 99% better) fairness in the homogeneous multiprocessor system and much better (more than 60% better) in the HMP system.