Using Centralized I/O Scheduling Service (CISS) to Improve Cloud Object Storage Performance

Load imbalance reduces performance of cloud object storage with restrained system utilization and cost performance. The state of the art is to model load status by probing data nodes for decentralized scheduling (e.g. C3). This paper exploits the potential of centralized scheduling. We design a centralized I/O scheduling service (CISS), mainly concerning availability (high-performance and fault tolerance). For high-performance, it uses several techniques, delivering throughput of scheduling decision making with 3 million/s. The key is that load status can be jointly learned with scheduling to cut mutual overhead. First, it uses the basic scheduling operation unit (BSOU) to combine scheduling and learning. Second, scheduling requests are packed into BSOU stream. Third, scheduling decisions are computed in sequence at server-side. For fault tolerance, CISS is developed with a stateless primary-backup model. We implement a distributed object storage prototype to evaluate CISS. Experiments show that CISS can deliver similar utilization and performance to C3. Compared with C3, CISS is better at reducing tail latency (up to 37.8% reduction of maximum latency). Moreover, CISS can quickly eliminate performance fluctuation caused by the statelessness of fault tolerance strategy, which is quite acceptable.