CMU is working toward a new approach to providing storage quality-of-service (QoS) in distributed environments supporting multiple services with time-varying workloads. Today's shared storage, whether implementing QoS or not, allows different services' workloads to mix without consideration of the large efficiency swings caused by unpredictable interference. Ideally, however, each service would see full efficiency within the fraction of the I/O system's time allocated to it, which would increase effective utilization and enable practical storage QoS control. We are formulating and beginning exploration of a resource management architecture in which QoS control is layered atop robust performance insulation and equipped with algorithms for dataset assignment and slack exploitation. The eventual outcomes seeded by this project will include enabling the storage QoS critically needed for emerging virtualized and cloud computing environments, enhancing education at CMU and elsewhere by providing insights taught in our storage systems and distributed systems classes, and being integrated into a deployed instance used by scientists sharing a cluster for their work.
