Modeling and design of continuous media storage architectures
D. Venkatesh, Thomas D. C. Little · 1998
The design of large-scale continuous media (CM) storage servers must balance the conflicting goals of achieving rapid response times and high server utilization. This dissertation addresses these issues in the context of disk-based server architectures and proposes a server design approach that scales across a spectrum of application bandwidth requirements. A data streaming model is considered to evaluate latency/bandwidth tradeoffs in disk systems. The model is subsequently extended to consider disks with Zone Bit Recording (ZBR) to exploit disk geometry to trade capacity for bandwidth by an optimal grouping of disk zones. Application of the design technique is shown to reduce disk counts compared with traditional data placement for several capacity/bandwidth requirements. A server cost-performance model is developed to determine disk parameters that yield the minimum cost per unit bandwidth for a given stream bandwidth, device capacity, and memory cost. This work demonstrates the inefficiencies of building server architectures based on streaming data for low bandwidth applications. The cost model also illustrates the need to constrain the number of drives in a disk array to prevent inefficient server operation. The server cost and disk streaming models form the basis of a design procedure that can be used to create large-scale CM servers given data describing user behavior, system component costs, media types and the desired levels of interaction. These results are applicable to the design of large-scale CM servers and are supported by extensive analysis, simulation, and measurement. A simulation study based on a workload derived from observations of user behavior in movie theaters and object size distributions based on data from the Internet Movie Database is used to demonstrate the design procedure.