Design of the Rio (randomized I/o) Storage Server

We describe the design and implementation of the RIO (Randomized I/O) multimedia object server which manages a set of parallel disks and supports real-time throughput and delay guarantees. This storage subsystem was implemented as part of a multimedia information server which supports multiple concurrent applications such as video on demand and 3D interactive virtual worlds. We discuss the principal issues and innovations involved in the design and implementation of the RIO storage system, and present experimental performance results measured on our prototype. A multimedia data server must be ready to handle a variety of realtime object types (video, audio, 3D interactive virtual worlds, etc.) as well as non realtime workload. Achieving simultaneously (1) high utilization and (2) low latency with a high degree of certainty is the challenge. Our prototype system provides a statistical guarantee of quality of service. Our experimental results shows that it is possible to achieve a very small probability of missing a deadline (less than 10 ?6), with a relatively high disk utilization (70% to 99%, in terms of fraction of the maximum disk capacity) , together with relatively small deadlines (on the order of 0.5 sec. to 1.5 sec), using contemporary disks. To achieve guaranteed low delay with high utilization requires good short term load balancing to keep disk queue length distributions from having long tails. For workloads with predictable access patterns it is possible to try to exploit this predictability to gain the desired eeect. However this is an approach with limited applicability and quickly becomes untenable as the workload becomes more diverse. Our approach is to randomize the physical allocation of disk blocks across the system's multiple disks. At the level of physical disk block access this turns all workloads into the same uniformly random access pattern and thus gives us one problem to deal with. Since each disk is equally likely to be the target of each disk access, this approach provides long range load balance. However short term statistical variations can result in short term imbalances in disk queues which in turn imply either (1) the guaranteed maximum latency must be increased or (2) a higher probability of missing the deadline. Our approach to counteracting this problem is to introduce limited redundancy by replicating (at random) some fraction of the data blocks. This data redundancy gives the disk scheduler some exibility in scheduling disk block reads, allowing the system to reduce …