Research On Topology Reconfiguration Algorithms For Many-Core Processors With Core-Level Redundancy Mechanism

As semiconductor technology entered into the nanometer age, and due to restricting factorssuch as the power consumption, the interconnect delay and the design complexity, the chipdesigning technology has shifted from the traditional mode of highly complex uniprocessors tothat of many-core processors, which is characterized by integrating a great amount of relativelysimpler processor cores on a single silicon die. It has been proliferated into the many-core era.However, with the ever-increasing risk of cores failure, which is caused either by manufacturingdefects and process variations, or by the ever-augmenting of core quantities on a single chip, thechip yields are shrinking. Therefore, how to improve the yields of processors has become a hotissue in industry and academia.The core-level redundancy technology is an effective way to improve the yields of many-coreprocessors. The existing row rippling column stealing (RRCS) topology reconfiguration algorithm,which bases on the idea of hierarchical optimization, divides the overall optimization problem intolocal optimization problems by the fault core, and achieves the overall optimal solution throughsearching the local optimal solution. However, the local neighborhood searching of the RRCSalgorithm is unidirectional, easily leading to the worse suboptimal solutions or the serious chaincolumn stealing operations that the former unit sequentially occupies the optimal solution of nextunit. Facing this situation, an optimized RRCS algorithm is proposed to better the local solutionand avoid the chain operations by using local neighborhood bidirectional searching. Experimentsshow that: the fault cores on physical topology is more, the optimized algorithm is better.In order to meet the applicable requirements of the many-core system in safety-critical area, areliability-centred topology reconfiguration simulated annealing algorithm for many-coreprocessors with core-level redundancy mechanism is presented. The algorithm firstly establishes areliability model for many-core processors with2D-Mesh NoC, then resorts to Monte-Carlo method to simulates their reliability by daggers sampling, finally captures their optimalreconfiguration topology by the simulated annealing optimization technique. Comparativeexperiments show that: the actual physical topology imposes non-prominent influence on theoverall reliability of many-core processors when their2D-Mesh network physical-links reliabilityis not too low. Therefore, the phenomenon indicates it only need to tackle the cores reliabilityimpact caused by the clustering effect of cores failure. The experiment also reveals the newalgorithm can improve the reliability of many-core processors significantly when the number ofsurviving cores are more.In this paper, we attempt to solve the problem of the core-level redundancy topologyreconfiguration, which is an instance of NP-Complete problem, by using the approximation andheuristic methods. The obtained results can provide some references or supports to tackle thetopology reconfiguration of many-core processors considering the performance and reliability.