CloudMix: Generating Diverse and Reducible Workloads for Cloud Systems

The prosperity of cloud computing offers commoninfrastructures to a wide range of applications. Understandingthese applications' workload behaviors is the premise of designing,managing, and optimizing cloud systems. Considering theheterogeneity and diversity of cloud workloads, for the sake offairness, cloud benchmarks must be able to accurately replicatetheir behaviors in cloud systems, including both the usages ofcloud resources and the micro-architectural behaviors beyondthe virtualization layer. Furthermore, workloads spanning longdurations are usually required to achieve representativeness inevaluation. Hence the more challenging issue is to significantlyreduce the evaluation duration while still preserving their workloadcharacteristics.This paper presents our efforts towards generating cloudworkloads of diverse behaviors and reducible durations. Ourbenchmark tool, CloudMix, employs a repository of reducibleworkload blocks (RWBs) as the high level abstraction of workloadbehaviors, including usages of the two most important cloud resources(CPU and memory) and their pairing micro-architecturaloperations. CloudMix further introduces an efficient methodologyto combine RWBs to synthesize and replicate diverse cloudworkloads in real-world traces. The effectiveness of CloudMixis demonstrated by generating a variety of reducible workloadsaccording to a Google cluster trace and by applying theseworkloads in job scheduling optimization on Hadoop YARN.The evaluation results show: (i) when the workload durationsare reduced by 100 times, the replication errors of workloadbehaviors are smaller than 2.08%; (ii) when providing fastevaluations (workload durations are reduced by 10 to 100 times)to recommend the optimal setting in YARN job scheduling,the performance degradation in the recommended setting isjust 0.69% compared to that of the actual optimal setting.CloudMix is publicly available from the projec.