User Modeling of Parallel Workloads

The goal of workload modeling is to simulate the expected workload, accurately enough to enable making correct design and administrative decisions. Several statistical features of production parallel computer workloads, which are not embodied in current models, have been identified. Their practical importance is demonstrated by two new kinds of schedulers – a key component in determining the overall performance of a parallel computer. The first is adaptive scheduling, which takes advantages of the locality of sampling and known cycles in parallel workloads, and achieves an average improvement of 10% in performance and 35% in stability for the tested production workloads. The second is shortest-jobbackfill-first scheduling, which relies on runtime prediction, done by analyzing user and session histories. These schedulers cannot be correctly evaluated by existing workload models, and we argue that the correct approach for future workload models (as well as on-line algorithms) is user and session-based modeling, instead of modeling jobs directly as done today. As the basis for such a model, we use PCA to provide variable sets which explain over 80% of the variance between users and sessions, and clustering to identify five stable session clusters and four stable user clusters. We then model the distributions of the arrival and activity patterns of both users and sessions, including a complete analysis of their dependencies and temporal structure. The model is based on logs from seven different parallel supercomputers, spanning over 87 months, analyzed together to ensure that results are location and architecture-neutral.