Reward Scheduling for QoS in Cloud Applications

We present a novel QoS scheduling algorithm for multi-tiered storage servers made up of hard disks and SSDs. The work is motivated by the difference in access times for a workload as its SSD hit ratio changes. Our scheme is designed to reward clients according to their runtime behavior, while honoring their static QoS settings including shares (or weights), reservations and limits. A model based on entitlements is developed to describe the reward allocation policy (RAP). Simulation results show the advantages of RAP over conventional proportional share allocation in adapting to dynamically varying workloads. The proposed algorithm allows the client to directly reap the benefits of application performance tuning as if the client is on a dedicated system, addressing a key complaint of clients when moving to a shared infrastructure.