Scale-Space Filtering for Workload Analysis and Forecast

Dynamic resource provisioning poses a major challenge for infrastructure providers because it is necessary to both forecast resource consumption and react to recent surges on demand for maintaining a tradeoff between quality of service and cost. However, approaches to workload analysis and forecast are affected due to noise in observed data, specially in forecast models. Moreover, most studies do not consider different prediction horizons may be necessary in order to take action before an SLA violation occurs. This paper presents an approach based in the scale-space theory to assist the dynamic resource provisioning. This theory is capable of eliminating the presence of irrelevant information from a signal that can potentially induce wrong or late decision making. In order to evaluate our approach, some experiments are presented considering both reactive and proactive strategies. The results confirm that our approach improves the workload analysis and forecast.