Control flow behavior of cloud workloads

With massive amounts of information on the web, cloud applications are rapidly emerging as one of the main-stream domains in modern computing, yet very little is known about their behavior. To our knowledge, this paper presents the first detailed study of control flow behavior in cloud workloads. We characterize branch predictability behavior of cloud and big data benchmarks, and compare against those of widely known CPU workloads based on profiling and simulation. Our in-depth branch analysis of workloads present striking differences in terms of higher prevalence of indirect branches, larger offsets in branch targets, abundance of multi-target branches and low BTB hit-rates. We identify performance bottlenecks involving branch predictability and provide suggestions that can be incorporated in future datacenter oriented processor designs. We perform Principal Component Analysis and clustering techniques to understand similarity/dissimilarity between cloud and CPU workloads.