Analysis of Mixed Workloads from Shared Cloud Infrastructure

Modern computing environments such as clouds, grids or HPC clusters are both complex and costly installations. Therefore, it has always been a major challenge to utilize them properly. Workload scheduling is a critical process in every production system with an unwanted potential to hamper overall performance if the given scheduler is not adequate or properly configured. Therefore, researchers as well as system administrators are frequently using historic workload traces to model/analyze the behavior of real systems in order to improve existing scheduling approaches. In this work we provide such real-life workload traces from the CERIT-SC system. Importantly, our traces describe a “mixed” workload consisting of both cloud VMs and grid jobs executed over a shared computing infrastructure. Provided workloads represent an interesting scheduling problem. First, these mixed workloads involving both “grid jobs” and cloud VMs increase the complexity of required (co)scheduling necessary to efficiently use the underlying physical infrastructure. Second, we also provide a detailed description of the setup of the system, its operational constraints and unresolved issues, putting the observed workloads into a broader context. Last but not least, the workloads are made freely available to the scientific community allowing for further independent research and analysis.