High-speed DFG-level SEU vulnerability analysis for applying selective TMR to resource-constrained CGRA
High-speed DFG-level SEU vulnerability analysis for applying selective TMR to resource-constrained CGRA
复制标题
DOI：
10.1109/isqed.2013.6523663
复制
发表时间：
2013-03
期刊：
影响因子：
--
通讯作者：
Takashi Imagawa;Hiroshi Tsutsui;H. Ochi;Takashi Sato
中科院分区：
文献类型：
--
作者：
Takashi Imagawa;Hiroshi Tsutsui;H. Ochi;Takashi Sato
研究了一种抗单事件翻转(SEU)的高性价比选择性三模冗余(Selective TMR)的实现方法。这种方法使我们能够最小化在资源受限的粗粒度可重构体系结构(CGRA)上实现的目标应用电路的脆弱性。该方法的关键是确定目标应用程序数据流图中易受攻击节点的评价函数。由于节点故障对主输出的影响取决于节点的扇区和扇区以及节点本身，本文提出了一种反映节点扇入/扇区运行情况的增强评价函数。本文还改进了评价函数中权重向量的计算方法，对节点的脆弱性假设为指数分布而不是线性分布。为了得到通用的权重向量，我们建议求解从多个样本应用中获得的级联线性方程，而不是为了应用而对权重向量进行平均。该方法利用广义逆矩阵解该方程，只需不到10秒的时间即可得到合理的优先级，比穷尽搜索最优解耗时15小时以上的速度快得多。本文还比较了评估函数中使用的特征的贡献，这将对设计具有可靠性意识的CGRA体系结构和综合工具具有一定的指导意义。
In this paper, we investigate a method to achieve cost-effective selective triple modular redundancy (selective TMR) against single event upset (SEU). This method enables us to minimize the vulnerability of the target application circuit implemented on a resource-constrained coarse-grained reconfigurable architecture (CGRA). The key of the proposed method is the evaluation function to determine the vulnerable node in the data flow graph (DFG) of the target application. Since the influence of the fault in a node to the primary outputs depends on its fains and fanouts as well as the node itself, this paper proposes an enhanced evaluation function that reflects the operation of fanins/fanouts of a node. This paper also improves the method to derive weight vector which is used in the evaluation function, by assuming exponential distribution instead of linear distribution for the vulnerability of nodes. To derive a generic weight vector, we propose to solve a concatenated linear equations obtained from multiple sample applications, instead of averaging the weight vectors for applications. Using generalized inverse matrix to solve the equation, the proposed method takes less than ten seconds to extract a reasonable priority for selective TMR, which is extremely faster than the exhaustive exploration for the optimal solution that takes more than 15 hours. This paper also compares the contributions of the features use in the evaluation function, which would be insightful for designing reliability-aware CGRA architecture and synthesis tools.