Workload modeling for parallel computers

Abstract

The availability of good workload models is essential for the design and analysis of parallel computer systems. A workload model can be applied directly in an experimental or simulation environment to verify new scheduling policies or strategies. Moreover, it can be used for extrapolating and predicting future workload conditions. In this work, we focus on the workload modeling for parallel computers. To this end, we start with an examination of the overall features of the available workloads. Here, we find a strong sequential dependency in the submission series of computational jobs. Next, a new approach using Markov chains is proposed that is capable of describing the temporal dependency. Second, we analyze the missing attributes in some workloads. Our results show that the missing information can be still recovered when the relevant model is trained from other complete data set. Based on the results of overall workload analysis, we begin to inspect the workload characteristics based on particular user-level features. That is, we analyze in detail how the individual users use parallel computers. In particular, we cluster the users into several manageable groups, while each of these groups has distinct features. These different groups provide a clear explanation for the global characteristics of workloads. Afterwards, we examine the user feedbacks and present a novel method to identify them. These evidences indicate that some users have an adaptive tendency and a complete workload model should not ignore the users' feedbacks. The work ends with a brief conclusion on the discussed modeling aspects and gives an outlook on future work.