Design of a fault tolerant real-time storage system for multimedia applications

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 statistical delay guarantees. The original implementation, on an SMP (Sun Microsystems E4000), has been ported to a cluster of PCs. The new version uses a similar philosophy with respect to allocation of data storage and scheduling of requests. However the parallel "shared nothing" architecture raises all the issues of latency, scalability, etc., that are familiar in parallel and distributed systems. Our requirements are particularly demanding since we expect to support a variety of multimedia applications with varying workload requirements and data reference patterns. 3D interactive applications, video, hypermedia and interactive scientific visualization of terabyte size data sets are among the applications we support. This paper details the design issues we addressed and the results of detailed simulation studies which dictated the design choices made.