On the Selection of Models for Runtime Prediction of System Resources

Applications and services delivered through large Internet Data Centersare now feasible thanks to network and server improvement, but also to virtualization,dynamic allocation of resources and dynamic migrations. The large numberof servers and resources involved in these systems requires autonomic managementstrategies because no amount of human administrators would be capable of cloningand migrating virtual machines in time, as well as re-distributing or re-mapping theunderlying hardware. At the basis of most autonomic management decisions, thereis the need of evaluating own global behavior and change it when the evaluationindicates that they are not accomplishing what they were intended to do or some relevantanomalies are occurring. Decisions algorithms have to satisfy different timescales constraints. In this chapter we are interested to short-term contexts whereruntime prediction models work on the basis of time series coming from samples ofmonitored system resources, such as disk, CPU and network utilization. In similarenvironments, we have to address two main issues. First, original time series areaffected by limited predictability because measurements are characterized by noisesdue to system instability, variable offered load, heavy-tailed distributions, hardwareand software interactions. Moreover, there is no existing criteria that can help us tochoose a suitable prediction model and related parameters with the purpose of guaranteeingan adequate prediction quality. In this chapter, we evaluate the impact thatdifferent choices on prediction models have on different time series, and we suggesthow to treat input data and whether it is convenient to choose the parameters of aprediction model in a static or dynamic way. Our conclusions are supported by alarge set of analyses on realistic and synthetic data traces.