高级检索
欧阳一鸣, 笪健, 李建华, 梁华国, 黄正峰, 杜高明. NoC中基于路径多样性的交叉开关细粒度容错设计[J]. 计算机辅助设计与图形学学报, 2017, 29(1): 180-188,210.
引用本文: 欧阳一鸣, 笪健, 李建华, 梁华国, 黄正峰, 杜高明. NoC中基于路径多样性的交叉开关细粒度容错设计[J]. 计算机辅助设计与图形学学报, 2017, 29(1): 180-188,210.
Ouyang Yiming, Da Jian, Li Jianhua, Liang Huaguo, Huang Zhengfeng, Du Gaoming. A Fine-Grained Fault-Tolerant Design of Crossbar Based on Path Diversity in Network-on-Chip[J]. Journal of Computer-Aided Design & Computer Graphics, 2017, 29(1): 180-188,210.
Citation: Ouyang Yiming, Da Jian, Li Jianhua, Liang Huaguo, Huang Zhengfeng, Du Gaoming. A Fine-Grained Fault-Tolerant Design of Crossbar Based on Path Diversity in Network-on-Chip[J]. Journal of Computer-Aided Design & Computer Graphics, 2017, 29(1): 180-188,210.

NoC中基于路径多样性的交叉开关细粒度容错设计

A Fine-Grained Fault-Tolerant Design of Crossbar Based on Path Diversity in Network-on-Chip

  • 摘要: 在片上网络中,交叉开关是路由器的重要组成部分,当其出现故障会导致数据无法传到下一跳,易造成资源浪费,影响整个网络的性能.对此,通过详细考虑交叉开关中数据分配器、内部链路以及数据选择器的故障情况,充分利用交叉开关的可用资源,提出交叉开关细粒度容错设计方案,实现容错路径的多样性.当数据到达数据分配器时,查看故障信息表,若存在故障则顺序选择一个无故障的数据分配器传输数据;当数据经过数据分配器传至内部链路时,若存在链路故障,同样查找故障信息表快速选择合适的容错链路进行传输;最后数据到达数据选择器,若此时数据选择器出现故障,数据则经过额外的数据选择器到达输出端口.实验结果表明,故障率逐渐增大的时候,该方案比已有方法平均延时降低了6%~36%,吞吐率提高了6.7%~27.7%,且容错效果不会随着故障率增大而显著下降.

     

    Abstract: Crossbar is an important part of the router in network-on-chip, when it breaks down, the data can't be transmitted to the next hop which will cause waste of resources and influence the performance of entire network. Considering faults of crossbar internal component including Demultiplexers, internal links and Multiplexers, in order to make full use of available resources of crossbar, this article proposes a fine-grained crossbar fault-tolerant design which can achieve the diversity of fault-tolerant path. Firstly, the data arrive to Demultiplexer, seeing the fault information table, and then choosing a trouble-free Demultiplexer orderly for data transmitting when there is fault. Secondly, the data arrive to internal links through Demultiplexer, choosing a suitable fault-tolerant link from the fault information table in the same way. Finally, the data arrive to Multiplexer. When faults occur, the data can be transmitted to the output port by the extra Multiplexer. Experimental results show that under the condition of the failure rate increase gradually, compared to comparative methods, our scheme can reduce average network latency by 6%~36%, increase throughput rate by 6.7%~27.7%, and the fault-tolerant effect will not decrease significantly with increasing failure rate.

     

/

返回文章
返回