Disk Subsystem Performance and Reliability Enhancements Through the Use of Carnival Mirrors

Performance and reliability modeling of computer systems continues to be an area of active research. Analytic and simulation modeling of computer systems and subsystems provides a cost effective means of system design, system expansion, and system tuning. Recent breakthroughs in computational power, coupled with the continuing development of the performance and reliability modeling field, now provide us with an opportunity to begin to answer design questions previously considered far out of reach. Traditional computer systems contained only single-unit disk subsystems. Many current computer systems have multiple-unit disks that are treated as a single logical unit. These multiple disk systems provide many additional opportunities for performance and reliability enhancement. We develop a new mirrored disk protocol that combines disk cylinder rearrangement with workload considerations to yield increased performance in the form of reduced expected seek distance for a mirrored disk subsystem. The new protocol calls for rearrangement of the cylinders on each disk according to non-identical permutations, in order to match characteristics of a measured workload. We term this approach carnival mirrors . We extend single disk cylinder rearrangement analysis by developing three equivalent models of the carnival mirror system: a Markov model, a recursive model, and a simulation model. We find that, subject to computational space and time constraints, any of the three models can be used as an estimator of expected seek distance within a simulated annealing routine that searches for optimal permutations of cylinders on the disks. Using this approach, we see as much as 70% improvement over a standard mirrored disk. As computational resources become more powerful and better optimization tech­niques for searching large solution spaces are developed, implementation of carnival mirrors becomes more feasible