软错误率变动对检查点机制的影响

发布时间：2018-09-07 21:40

【摘要】：随着集成电路工艺进入纳米时代,芯片正面临着软错误的威胁.除了软错误在数量上的威胁外,处理器还面临着由于工艺变动性和电压、温度、位置等工作环境变动而导致的错误率变动性的威胁,即系统的错误率不会一直稳定而是随着时间发生变化.检查点是系统容错的主要机制,它的开销和检查点的间隔密切相关,目前检查点间隔的确定大多是基于恒定错误率的.但是在软错误变动的情况下,自适应检查点的方法比固定方法更能够显著降低检查点开销,它通过预测系统的错误率来确保系统的检查点间隔始终与最优状态接近.但是自适应检查点所能获得的性能改善与错误率变动的具体程度相关.因此本文研究软错误率变动的形式和幅度如何影响检查点的开销.该文开展了如下研究:基于温度、电压、位置等因素对软错误影响的原理,建立了一个包含变动幅度、持续时间等参数的错误率变动的模型;基于错误率变动模型,模拟了在理想情况下自适应检查点机制能够获得的性能改善;提出了一种基于错误历史预测错误率的方法,从而验证了在实际情况下自适应检查点能够达到的效果.实验结果表明,变动的幅度在3倍以上且持续时间在12.5%以上时,该文方法就能获得实际上的性能改善.
[Abstract]:With the IC technology entering the nanometer age, the chip is facing the threat of soft error. In addition to the quantitative threat of soft errors, processors also face the threat of varying error rates due to process variability and changes in working conditions such as voltage, temperature, location, etc. That is, the error rate of the system does not remain constant but changes over time. Checkpoint is the main fault-tolerant mechanism in the system, and its overhead is closely related to the interval between checkpoints. At present, the determination of checkpoint interval is mostly based on constant error rate. However, in the case of soft error variation, the adaptive checkpoint method can significantly reduce the checkpoint overhead than the fixed method. It can predict the error rate of the system to ensure that the checkpoint interval of the system is always close to the optimal state. However, the performance improvement achieved by adaptive checkpoint is related to the specific degree of error rate variation. Therefore, this paper studies how the change of soft error rate affects the cost of checkpoint. Based on the influence of temperature, voltage, position and other factors on soft error, a model of error rate variation including amplitude and duration is established, and the error rate model is used to analyze the influence of temperature, voltage and position on soft error. The performance improvement of adaptive checkpoint mechanism under ideal condition is simulated, and an error rate prediction method based on error history is proposed, which verifies the effect of adaptive checkpoint in practice. The experimental results show that when the range of variation is more than 3 times and the duration is more than 12.5%, the performance of this method can be improved in practice.
【作者单位】：国防科学技术大学计算机学院;
【基金】：国家自然科学基金(61402504,61033008,61272145,61103080) 国家“八六三”高技术研究发展计划项目子课题(2012AA012706) 高等学校博士学科点专项科研基金优先发展领域课题(20124307130004)资助~~
【分类号】：TN407

【相似文献】