Performance evaluation of mass storage systems for scientific databases

Abstract

Mass storage systems for computers are the solution to economic storage of vast volumes of data. These systems evolved from the traditional tape libraries manned by operating personnel and the automation of the storage and retrieval function has led to significant improvement in performance. But in contrast to traditional computer systems, little work has been done to characterize performance in terms of the design parameters. The design and performance analysis of mass storage systems is complicated due to several reasons. A major reason for the complexity is the time lags that may occur in retrieving parts of the information meant for the same query. The usual queuing models used for analyzing disk performance are not directly applicable, because there is greater scope for working in parallel in mass storage systems such as robotic libraries, which help mitigate these shortcomings. In this note, robotic libraries are modeled as queueing systems and explicit results related to performance are obtained. The physical model corresponds to a mass storage system, where the information is stored in cassettes, which are retrieved by robots to be read using one or two read heads. The results pertain to the effect of file splitting on cassettes, and optimal configuration and control of robots that perform the retrieval and storage functions.