Investigating Core-Level N-Modular Redundancy in Multiprocessors
Aleksandar Simevski, Rolf Kraemer, Milos D. Krstic · 2014
When the limits of exploiting parallelism are reached, multiprocessors are not efficient since many of the processing elements (cores) are idle. Several proposed solutions use this idleness in order to improve the fault tolerance of the multiprocessor, with relatively small performance and/or cost overheads. In this paper, we propose a dynamical fault tolerant scheme in which one or more groups of multiprocessor cores are arranged to form tightly-synchronized N-Modular Redundant (NMR) sub-systems. These NMR groups could be formed dynamically, with an arbitrary number of cores, according to the application requirements of fault tolerance and performance. Moreover, the cores which are not engaged in an NMR group could continue to work independently. Tight synchronization and voting in each clock cycle is provided by special, programmable NMR voters. We make a vast number of experiments in order to investigate the reliability of the proposed scheme. The results led to closed-form expressions which show whether a given NMR group of N cores will be able to withstand the assumed fault rate during the projected mission time.