Analysis of Multiprocessor Control Organizations with Partial Program Memory Replication

Abstract

In certain real-time multiprocessor controllers the program memory must be at least partly replicated to reduce queuing delays and meet throughput requirements. In these organizations the program memory is a common pool consisting of K sections servicing m CPU's. The jth section is of size Xi and is replicated ni times (ni ≤ m); j = 1,...-, K. Dynamic programming is used to find optimal values of the number of replications ni relative to given fixed values of m, K, and Xi. This solution satisfies throughput requirements at minimum cost, relative to the given m, K, and Xi. The process is repeated for several values of the last three parameters, using a systematic search procedure described herein, until reaching the first local minimum. This usually provides a satisfactory suboptimal solution. Partial replication (n, ≤ m) is economically attractive in a wide class of real-time control systems in which the CPU's access a relatively small portion of the program store most of the time. Quantitative examples indicate that partial replication may cause 30 to 40 percent cost reduction compared to full replication.