Optimal replication of parallel-read, sequential-write systems

Frequently used computer elements that can be written by at most one process at a time constitute important bottlenecks in multiprocessor system operation, particularly when such elements are accessible only serially. Hardware devices, data files, system tables and critical sections in general may be examples of such elements. One common way to relieve this congestion is to provide several copies of the element, which can then be read (used) in parallel. However, the requirement that writing (changing) remain sequential means that writing times increase with the number of copies provided. The optimization question in this trade-off is the main concern of this paper. A probability model of such a system is formulated with the objective of obtaining read-rate capacities as a function of read/write loads and the number of copies provided. The above optimization problem is expressed in terms of these results and then solved. In particular, it is shown how to select the number of copies that maximizes the read-rate capacity for given system parameters. Two distinct operating regimes, based on how interrupted read operations are restarted, are analyzed.