Workload modeling and performance evaluation in parallel systems
Tran Ngoc Minh · Leiden Repository (Leiden University) · 2012
Scheduling plays a significant role in producing good performance for clusters and grids. Smart scheduling policies in these systems are essential to enable efficient resource allocation mechanisms. One of the key factors that have a strong effect on scheduling is the workload. This workload problem is associated with four research topics to obtain an effective scheduler, namely workload characterisation, workload modeling, performance evaluation and prediction, and scheduling design. Workload data collected from real systems are the best source for improving our knowledge about performance issues of clusters and grids. Observed features of these workloads are precious sources of clues, which can be utilized to enhance scheduling. To this end, several long-term parallel and grid workloads have been collected and this thesis used these real workloads in the study of workload characterisation, workload modeling, per formance evaluation and prediction. Our research resulted in many workload modeling tools, a performance predictor and several useful clues that are essential to develop efficient cluster and grid schedulers.