Filtering multivariate workload non-conformance in shared IT-infrastructures

Virtualized data centers are hosting virtual machines (VMs) running enterprise applications with time varying resource demand behavior (workload) jointly on the same physical servers in order to increase server utilization. To avoid server overload situations, a data center operator needs to decide which VMs should be assigned to a physical server for a given period of time. As assignment decisions are based on historical workload traces, deviations from forecasted demands potentially require the adaptation of VM assignments. However, VM live migrations are inherently overhead afflicted control operations. While it is mandatory to anticipate overload situations in order to trigger VM migrations in a proactive manner, unnecessary VM migrations may have negative impact on the underlying computing infrastructure. Hence, an autonomic controller should accurately predict situations where the aggregated workload of a set of collocated VMs will hit the capacity limit of a server without requiring manual adjustments of control model parameters. In this paper we propose an automated, non-parametric approach for proactive filtering of multivariate workload behavior. We learn an orthonormal projection from historical workload traces and extract a set of key metrics that concisely describe relevant developments in the joint workload behavior of physical servers. A geometric interpretation, in combination with simple short term forecasting techniques allows for reliable decision making. Based on a set of real world workload traces we conduct numerical experiments that validates its superiority and predictive capabilities over simple threshold based approaches.