Exploring Potential for Resource Request Right-Sizing via Estimation and Container Migration in Apache Mesos

Both commercial clouds and academic campus clusters suffer from low resource utilization and long wait times as the resource estimates for jobs, provided by users, is often inaccurate. Incorrect resource estimation poses challenges in the overall cluster and cloud management. Under allocation can cause significant slowdown or termination of applications. Over-allocation of resources for applications causes increased wait times for pending tasks in the queue, reduced throughput, and underutilization of the cluster. For end users that pay for resource allocations, incorrect estimation of resources (CPU, Memory, etc.) that are needed for each job can significantly increase the overall cost of running applications. Also, for academic cloud managers, resource fragmentation is unacceptable as they need to keep the utilization high to maximize the return on investment for the funding sponsors. We address the resource estimation problem for commercial and academic clouds/clusters that use the Apache Mesos resource management system. Our vision is a resource management system for Apache Mesos that can: (1) dynamically right-size the resources required for each application, thus improving overall utilization; and (2) incorporate migration of containerized jobs within the Mesos cluster.