A MAP Fitting Approach with Joint Approximation Oriented to the Dynamic Resource Provisioning in Shared Data Centres

Abstract

In shared data centres, accurate models of workloads are indispensable in the process of autonomic resource scheduling. Facing the problem of parameterizing the vast space of big MAPs in order to fit the real workload traces with time-varying characteristics, in this paper we propose a MAP fitting approach JAMC with joint approximation of the order moment and the lag correlation. Based on the state-of-the-art fitting method KPC, JAMC uses a similar divide and conquer approach to simplify the fitting problem and uses optimization to explore the best solution. Our experiments show that JAMC is simple and sufficient enough to effectively predict the behavior of the queueing systems, and the fitting time cost of a few minutes is acceptable for shared data center. Through the analysis of the sensitivity to the orders fitted, we deduce that it is not the case that the higher orders have better results. In the case of Bellcore Aug89, the appropriate fitted orders for the moments and autocorrelations should be respectively on a set of 10~20 and 10000~30000.