Storage system architectures for continuous media data

Abstract

Data storage systems are being called on to manage continuous media data types, such as digital audio and video. There is a demand by applications for “constrained-latency storage access” (CLSA) to such data: precisely scheduled delivery of data streams. We believe that anticipated quantitative improvements in processor and storage-device performance will not be sufficient for current data management architectures to meet CLSA requirements. The need for high-volume (but high-latency) storage devices, high-bandwidth access and predictable throughput rates mean that standard latency-masking techniques, such as buffering, are inadequate for the service demands of these applications. We examine the ways in which storage system architectures must change in order to provide CLSA on continuous media, taking into account operating system and network support as well as database management. Particular points we cover include

• changes in the form of requests and responses at the application-database and database-OS interfaces

• new kinds of abstractions and data independence that data mangement systems will need to supply, such as quality-of-service requests and mapping of domain events to OS events

• effects of CLSA demands on query optimization, planning and evaluation, including the need for accurate resource estimates and detailed schedules

• new information requirements for the database system, such as better characterizations of storage subsystem performance and application patterns.

We illustrate the problems and new demands of continuous media access with application requirements of an all-digital television production studio of the future. We conclude by presenting a storage architecture we are pursuing that we believe will support “scripted” CLSA to shared data.

This research is supported by NSF Grant IRI-9117008 and by funds from Tektronix, Inc. and the Oregon Advanced Computing Institute.