Workload Prediction for Runtime Resource Management
IEEE Nordic Circuits and System Conference (NorCAS 2017)
An intelligent resource manager is an essential part of platforms based on heterogeneous architectures. The resource manager should be able to accurately predict the future workload of the system at hand and take it into consideration for making decisions. In this paper, we study a large computer cluster and show that there exist patterns in the sequence of applications that each user runs over time, and that these patterns can be used for modeling and prediction of the applications that will be requested in the future. To this end, we develop a predictive technique based on the n-gram model. It is shown that, due to the varied nature of application sequences of different users, a universal model does not provide optimal results, and a customized model should be constructed for each user. The experimental results show that the straightforward methods have a prediction accuracy below 16% when assessed on a real-life data set. Our technique provides an accuracy improvement of more than 51% in comparison with the straightforward method.
[NUEP17] Mina Niknafs, Ivan Ukhov, Petru Eles, Zebo Peng, "Workload Prediction for Runtime Resource Management", IEEE Nordic Circuits and System Conference (NorCAS 2017)