Abstract

Data center workload modeling has become a necessity in recent years due to the emergence of large-scale applications and cloud data-stores, whose implementation remains largely unknown. Detailed knowledge of target workloads is critical in order to correctly provision performance, power and cost-optimized systems. In this work we aggregate previous work on data center workload modeling and perform a qualitative comparison based on the representativeness, accuracy and completeness of these designs. We categorize modeling techniques in two main approaches, in-breadth and in-depth, based on the way they address the modeling of the workload. The former models the behavior of a workload in specific system parts, while the latter traces a user request throughout its execution. Furthermore, we propose the early concept of a new design, which bridges the gap between these two approaches by combining some features from each one. Some first results on the request features and performance metrics of the generated workload based on this design appear promising as far as the accuracy of the model is concerned.