A MEASUREMENT-BASED MODEL TO PREDICT THE PERFORMANCE IMPACT OF SYSTEMMODIFICATIONS - A CASE-STUDY

This paper presents a performance case study of parallel jobs executin g in real multi-user workloads. The study is based on a measurement-ba sed model capable of predicting the completion time distribution of th e jobs executing under real workloads. The model constructed is also c apable of predicting the effects of system design changes on applicati on performance. The model is a finite-state, discrete-time Markov mode l with rewards and costs associated with each state. The Markov states are defined from real measurements and represent system/workload stat es in which the machine has operated. This paper places special emphas is on choosing the correct number of states to represent the workload measured. Specifically, the performance of computationally-bound, para llel applications executing in real workloads on an Alliant FX/80 is e valuated. The constructed model is used to evaluate scheduling policie s, the performance effects of multiprogramming overhead, and the scala bility of the Alliant FX/80 in real workloads. The model identifies a number of available scheduling policies which would improve the respon se time of parallel jobs. In addition, the model predicts that doublin g the number of processors in the current configuration would only imp rove response time for a typical parallel application by 25%. The mode l recommends a different processor configuration to more fully utilize extra processors, This paper also presents empirical results which va lidate the model created.