An Application Aware Approach to Scalable Online Placement of Data Center Workloads

. Data center operators strive for maximum resource utilization while satisfying tenant service level agreements however, they face major challenges as application workload types are diverse and tenants add, remove, and update their workloads sporadically to meet changing user demands. Currently, operators allocate workload VMs primarily in an application agnostic fashion, focusing on minimizing total resource usage. In this work, we first show that such allocations can be suboptimal and then present a new application aware approach that explicitly models resource preferences of individual workloads. Further, we propose a new logical application workload LAW abstraction to enable precomputation of the required relative positioning of an applications VMs and allocation of these VMs in a single atomic step, leading to online algorithms that are one order of magnitude faster than existing per VM placement solutions. We then develop a statistical extension of LAW to add flexibility in characterizing application requirements and to support prioritization of workloads. Using realistic workload traces and physical topologies, we evaluate our algorithms in a simulated larges scale data center setting, and demonstrate their performance advantages and potential tradeoffs versus existing solutions