Event Aware Workload Prediction: A Study Using Auction Events

Abstract

Workload bursts have become notorious for rendering numerous web information systems unavailable. While cloud computing has the potential to alleviate this problem by offering computing resources on an on-demand basis, important challenges remain in finding the right resource control strategies to scale resources cost-effectively and to overcome the initialization lag associated with resource acquisition. An effective strategy involves predicting workload demand in advance so that resources can be provisioned in a timely manner, but not all prediction approaches are made equal. We argue that while most existing approaches show promising results in predicting average workload, they fail to predict workload bursts that are inherently irregular. This paper formulates a new event-aware strategy to more effectively predict workload bursts by exploiting prior knowledge associated with scheduled events. We evaluate our approach by comparing it to state-of-the-art methods in workload prediction using real-world datasets from the online auction domain, and we show that event-aware prediction is superior to other approaches in terms of burst prediction accuracy.