| Hi-index | 0.00 |
Over the last decade, storage has been playing catch-up with the meteoric improvement in computation and communications infrastructure. However, this trend requires a change if future systems must be able to support emerging applications that deal with massive amounts of data. These high data-volume applications require that storage systems support not just the traditional quality-of-service (QoS) metrics like reliability, availability, etc. but also new metrics like real-time data-delivery, and low response-time while maintaining high system throughput. The development of conventional real-time systems has mainly focused on supporting the real-time paradigm for computational resources and usually assumes that all data resides in main memory. Real-time multimedia systems, however, manage large amounts of heterogeneous data that require to be placed on secondary storage. These include data that have real-time delivery constraints, traditional data that require best-effort service, as well as interactive data that require immediate service. In addition to satisfying computational real-time constraints, such systems must also support the real-time data storage and retrieval requirements of multimedia applications. This dissertation focuses on storage systems that can support the varied requirements of heterogeneous multimedia data. For providing the IO guarantees required by such data as well as improving the overall cost-performance metric of the storage system, we propose and evaluate two principal approaches. The first approach presents several methods for managing a traditional disk-only storage system to meet real-time delivery constraints as well as improving the response time of interactive operations. Unlike previous methods, these methods are developed based on accurate low-level profiling of storage device parameters. The second approach proposes new architectures for storage systems using MEMS-based storage, an emerging technology, and evaluates its value for real-time multimedia applications. This approach is a specific-case solution for the more general problem of managing heterogeneous data and storage. We propose rules-of-thumb for partitioning data types across storage devices and propose architectures and mechanisms for building MEMS-disk storage systems that can support real-time streaming data.