The data-centre whisperer: Relative attribute usage estimation for cloud servers

We show that the relative usage of the different attributes of a cloud server can be estimated under time-varying loads. We demonstrate the effectiveness of these estimators by determining how user requests for video-from a video server-affects its usage. Relative Attribute Usage (RAU) estimators are designed by (1) formulating a generative model for the server attributes; (2) using the fact that the load signal has compact support compared to non-idealities in the server's behaviour in the time-frequency domain; and (3) using power-weighting to refine the estimates. The resulting estimators have low complexity. This motivates their candidacy when attribute usage estimates are required for run-time outage diagnosis routines, a task which is commonly referred to as "data-centre whispering". We demonstrate the application of these estimators on a Cloudtestbed in three practical scenarios, when the server is under a (1) periodic, (2) step-increasing and (3) flash-crowd load.