Reproducible Partial-Load Experiments in Workload Colocation Analysis

Hardware concurrency is common in all contemporary computer systems. Efficient use of hardware resources requires parallel processing and sharing of hardware by multiple workloads. Striking a balance between the conflicting goals of keeping servers highly utilized and maintaining a predictable performance level requires an informed choice of performance isolation techniques. Despite a broad choice of resource isolation mechanisms in operating systems, such as pinning of workloads to disjoint sets of processors, little is known about their effects on overall system performance and power consumption, especially under partial load conditions common in practice. Performance and performance interference under partial processor load is analyzed only after the fact, based on historical data, rather than proactively tested. This dissertation contributes a systematic approach to experimental analysis of application performance under partial processor load and in workload colocation scenarios. We first present a software tool set called Showstopper, capable of achieving and sustaining a variety of partial processor load conditions. Based on arbitrary pre-existing computationally intensive workloads, Showstopper replays processor load traces using feedback control mechanisms to maintain the desired load. As opposed to..