Performance evaluation of hierarchical mass storage systems
Odysseas I. Pentakalos, Yelena Yesha, Daniel A. Menascé · 1996
One of the most interesting grand challenges that will drive high-performance computing and communications in the coming decade is the analysis of enormous amounts of satellite and simulation data to understand global effects such as weather front interactions and ocean-land-atmosphere interactions. These enormous amounts of data will impose tremendous demands on the underlying storage systems. To allow for the efficient storage and retrieval of such enormous amounts of data it is necessary to develop high-performance mass storage systems. Mass storage systems consist of a hierarchy of storage media arranged in such an order as to provide a fast access, high capacity, low cost storage system. The continuous increase in user demands at scientific computing research centers for storage space forces system administrators to procure additional storage devices without access to tools for justifying such procurement. Furthermore, the plethora of available storage devices in the market with different performance characteristics makes this capacity planning process even more difficult. This dissertation addresses this lack of tools for performance evaluation and capacity planning of hierarchical mass storage systems. Its main contributions are the development of queueing network models for both host attached and network attached device based mass storage systems. Two approximations were developed as components of this work. The first is an approximation of the performance of RAID devices in the setting of mass storage systems with multiple classes of requests. The second is an approximation of the simultaneous resource possession problem which occurs during tape to disk transfers in mass storage systems. The accuracy of both approximations was validated using process-based simulations. The third contribution of this dissertation is the development of Pythia--a tool for performance evaluation of hierarchical mass storage systems. Using the graphical user interface of Pythia, a user can easily describe the architecture of a mass storage system. The tool then automatically generates a queueing network model of the mass storage system and solves it using the modified multi-class approximate Mean Value Analysis algorithm with the approximations developed as part of this research.