基于随机网络演算的片上网络性能分析

发布时间：2019-04-25 06:48

【摘要】：性能分析是当今片上网络的研究热点之一,其中网络演算作为当今最有前途的网络性能分析工具,在片上网络领域中得到充分的应用。基于确定性网络演算的最差情形性能分析,由于其理论的出发点通常是最坏情况,所以得出的性能分析结果过于悲观,容易造成片上网络资源的浪费。基于随机网络演算的片上网络性能分析,能够更为准确的预测实际性能情况。本文以片上网络为研究对象,研究基于随机网络演算的性能分析建模方法。主要工作如下：1.提出基于服务期望的片上网络多条流冲突分析模型。传统的冲突分析模型仅仅根据业务流的路径确立冲突情况,忽略了流中数据包的影响,造成分析结果与实际性能情况存在一定差距。本文根据业务流的泊松到达过程,从业务流中数据包层面,计算出多条流冲突的可能,并利用随机网络演算知识建立冲突情况下每条业务流端到端的性能分析模型。仿真实验表明,相比Lu和Jing的冲突模型,在准确性方面至少得到18%以上的提升,验证了本文所提出模型的实用性和紧凑性。2.提出一种全局冗余TSV容错模型。已有的局部冗余TSV容错模型在提升芯片成品率方面的能力有限,需要一种更为高效的替代方案。本文提出的一种全局冗余TSV容错模型相比局部冗余TSV容错模型,在相同的冗余数目下,能够提升更高的芯片成品率,对比实验表明,在特定的冗余数目下,本文提出的模型相比局部冗余容错在成品率方面最高提升了29%。由于TSV故障率的概率性导致节点服务的不确定性,符合随机网络演算的建模思想,因此,本文利用随机网络演算建立了全局容错的节点服务模型。
[Abstract]:Performance analysis is one of the hot topics in the research of on-chip network, and network calculus, as the most promising tool for network performance analysis, has been fully applied in the field of on-chip network. Based on the worst-case performance analysis of deterministic network calculus, because the starting point of its theory is usually the worst case, the results of performance analysis are too pessimistic and easily lead to the waste of network resources on the chip. The performance analysis of on-chip network based on stochastic network calculus can predict the actual performance more accurately. In this paper, the modeling method of performance analysis based on stochastic network calculus is studied based on the on-chip network. The main work is as follows: 1. A multi-stream collision analysis model for on-chip network based on service expectation is proposed. The traditional conflict analysis model only establishes the conflict situation according to the path of the traffic flow, neglects the influence of the data packet in the flow, and results in a certain gap between the analysis result and the actual performance. In this paper, according to the Poisson arrival process of traffic flow, the possibility of multiple stream conflicts is calculated from the packet level in the traffic flow, and the end-to-end performance analysis model of each traffic flow in the case of conflict is established by using stochastic network calculus knowledge. Simulation results show that compared with Lu and Jing's conflict model, the accuracy of the proposed conflict model is improved by at least 18%, which verifies the practicability and compactness of the proposed model. 2. A global redundant TSV fault tolerant model is proposed. The existing partially redundant TSV fault-tolerant models have limited ability to improve the yield of chips, so a more efficient alternative is needed. A global redundant TSV fault-tolerant model proposed in this paper can improve the chip yield under the same number of redundancy compared with the partial redundant TSV fault-tolerant model. The experimental results show that under the given redundancy number, the chip yield can be improved more efficiently than that of the local redundant TSV fault-tolerant model. Compared with the local redundancy fault tolerance, the proposed model improves the yield by up to 29%. Because the probability of TSV failure rate leads to the uncertainty of node service, it accords with the modeling idea of stochastic network calculus. Therefore, this paper presents a global fault tolerant node service model based on stochastic network calculus.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN47

【相似文献】