The perceived effect of breakdown and repair on the performance of multiprocessor systems

It is argued that classical measures of computer system performance, for example mean response time, are inadequate in the context of fault-tolerant system design. Alternative, perception-based measures are proposed and theorems established describing their properties. Focus is directed upon the homogeneous M/M/m system in which total processor power is constrained by budget and processors are subject to failure and repair. A numerical technique for extracting both classical and perception-based measures from the associated two-dimensional Markov process is offered, along with bounds on time and space required for its execution. It is seen that the perception-based approach to system design can call for twice as many processors as the classical approach.