Abstract:
Ensuring high reliability and availability of virtualized NVMe storage systems is crucial for large-scale clouds. However, previous I/O virtualization systems only focus on improving I/O performance and ignore the above challenges. To this end, we propose Spool, a reliable NVMe virtualization system. Spool has three key advantages: (1) It diagnoses the device failure type and only replaces the NVMe devices with actual media errors. Other data link errors are handled through resetting the device controller, minimizing data loss due to unnecessary device replacement. (2) It ensures the consistency and correctness of the data when resetting the controller and upgrading the storage virtualization system. (3) It greatly reduces the restart time of the NVMe virtualization system. The quick restart eliminates complaints from tenants due to denial-of-service during a system upgrade and failure recovery. Our evaluation shows that Spool provides reliable storage services with performance loss smaller than 3%, and it reduces restart time by 91% when compared with SPDK.
