Stonehenge: a high-performance virtualized ip storage cluster with qos guarantees

Persistent information plays a critical role in enterprise and personal computing. Storage capacity requirement is reported to double each year. With the elevating gap between the increasing storage management cost and the decreasing hardware cost, storage manageability becomes more and more important. Storage storage management software and greatly simplifies storage management. However, current virtualization technique is only in capacity dimension and does not care about performance issues. In a shared storage environment, quality of service must be provided so that various clients can get guaranteed storage services. This dissertation presents a design, implementation and evaluation of a clustered storage system called Stonehenge, which provides multi-dimensional (bandwidth, capacity, latency) storage virtualization and run time service guarantees. To realize storage virtualization, we developed a virtual to physical mapping scheme which aims at maximizing the resource usage. Quality of service guarantee always translates to overprovision due to bursty traffic pattern. Besides a real-time scheduler to enforce the service level agreements, we also propose and evaluate a measurement-based feedback path to reflect the actual resource usage and maximize the virtual disks admission. Track based logging (Trail) is used to realize fast write caching, which provides high write performance with persistence warranty. It eliminates expensive NVRAM by using more cost-effective hard drives. Our performance study shows that Trail can reduce the synchronous write performance to data transfer time and fixed command processing overhead. Lastly, we report our experience in realizing the storage cluster on commodity hardwares and IP networks and their impacts on overall system performance.