Operation-Level Performance Control in the Object Store for Distributed Storage Systems

Storage QoS is getting more important to host big data and data intensive computing applications in the cloud. However, many distributed/file storage systems do not implement end-to-end performance control, and an abstraction layer of the storage device like the object storage device serves requests in a best effort basis. Thus the QoS granularity and accuracy seen in applications are rough and possibly fluctuating. In order to solve this problem, our proposed object store allows an upper layer of it to control operation-level performance, instead of only relying on raw level disk I/O scheduling. Our preliminary experiments show that the performance is well controlled by using our operation-level scheduling, as requested by multiple clients that would be launched by the upper layer.