Performance measurement and analysis of large filestores

Abstract

Performance measurements of two large time-sharing computer systems are presented, with emphasis on their disk filestores. Similarities of process behaviour are found in the measured systems and another system reported in the literature. Individual processes make i/o requests in sequences, or bursts. Burst lengths have a mean of two with a large variance; within a burst, file i/o requests are spatially sequential in intent and are temporally related. Characterizations of these behaviour patterns form the basis of a methodology for filestore evaluation and design. Descriptions of spatial and temporal load are abstracted from software traces without loss of any performance factor; these descriptions are inputs to a statistical model of the processes in the environment of the filestore. The filestore is represented by a simulation queuing model. The method specifies the inputs to the composite model and describes the calibration of outputs to match observable outputs. A model is built by this method, and validated for different loads. The model is used for three evaluation experiments. Disk request scheduling is not statistically significant; filestore layout and disk capacity are highly significant; disks with fast-access areas are shown to improve performance by taking advantage of spatial accessing patterns. The limits of performance of a novel filestore equipped with a cache store are explored to determine guidelines for this new design. Modest improvements resulting from this design are shown to produce a considerable improvement in overall system performance.