Restating the Case for Weighted-IPC Metrics to Evaluate Multiprogram Workload Performance

Weighted speedup is nowadays the most commonly used multiprogram workload performance metric. Weighted speedup is a weighted-IPC metric, i.e., the multiprogram IPC of each program is first weighted with its isolated IPC. Recently, Michaud questions the validity of weighted-IPC metrics by arguing that they are inconsistent and that weighted speedup favors unfairness [4]. Instead, he advocates using the arithmetic or harmonic mean of the raw IPC values of the programs in the multiprogram workload. We show that weighted-IPC metrics are not inconsistent, and that weighted speedup is fair in giving equal importance to each program. We argue that, in contrast to raw-IPC metrics, weighted-IPC metrics have a system-level meaning, and that raw-IPC metrics are affected by the inherent behavior of the programs. We also show that the choice of a metric may adversely affect the conclusions from an experiment. We suggest to use two weighted-IPC metrics - system throughput (STP) and average normalized turnaround time (ANTT) - for evaluating multiprogram workload performance, and to avoid raw-IPC metrics.