Using instruction result locality and re-execution to mitigate silent data corruptions

摘要

In this paper, a method to mitigate silent data corruptions (SDCs) is proposed. This paper, first, shows and characterizes instruction result locality based on several simulation results and next, proposes an architecture called instruction value history table (VHT) to detect SDCs. In the case of fault detection, extra instruction redundant execution is utilized to assure fault existence. If outcome of the new redundant execution is different from that of previous one, a fault occurred, otherwise the first execution will be correct. In order to correct any detected faults, third redundant execution of the instruction is performed. Having three values from three redundant instruction executions, makes the correction of the fault feasible. The main advantage of this method is to detect any error which is not detectable by traditional protection codes like parity and SEC-DED. In other words, this method detects SDCs or any multiple faults which are not detectable by protection codes. Various soft error injections have been applied on Alpha processor for several PARSEC benchmarks. Experimental results show that the method can detect up to 70% of injected SDCs. (C) 2016 Elsevier Ltd. All rights reserved.