Filtering observations to improve resource control in virtual computer and network systems

The Autonomic Resource Control Architecture (ARCA) we have designed in previous work is intended to provide elastic resource adaptation in virtual computer and network systems by automating management tasks in edge/branch virtual networks. Our objective is to reduce the time required to continuously estimate or anticipate the amount of resources that a system requires by considering its workload together with event data notified by external detectors. However, analyzing such amount of data can take a long time and thus delay management decisions and system adaptation. Moreover, too frequent changes in the amount of allocated resources can hassle the underlying controllers. In this paper we propose a method to resolve these problems by filtering input data items to reduce their rate while smoothing non-persistent spikes to reduce the volatility of resource allocations. We evaluate it by running an algorithmic model with a set of input data following a Pareto distribution, analyzing the accuracy of different variations of the proposed method. We found that the accuracy of the resulting allocations is improved by 13% to 30%, depending on the specific configuration of the proposed method.