Performance and availability evaluation of NUMA architectures

A new approximation technique is proposed for obtaining analytic estimates of the performance of computing systems exhibiting non-uniform memory access (NUMA) times, where components are subject to failure and repair. The technique uses a multi-chain mean-value analysis together with a service time scaling factor derived to correct for the non-exponential (deterministic) service times found in most service units of a NUMA architecture. The new technique is seen to provide excellent agreement with simulation results in predicting both mean network cycle time and contention at the deterministic system devices. Availability estimation is based on a machine-repair model with deterministic repair (replacement). In this case, exact analysis is available using techniques from Markov regenerative stochastic Petri nets.