Abstract: System area network is the important component of high performance computer and data center. The current system area network has the characteristics of larger-scale network, more internal links, more complex behavior of network traffic, higher sensitivity of applications on network performance and so on. Those problems cause that the computational complexity of the performance measurement is exponential growth of the traditional network tomo-
graphy. To address this problem, this paper proposes a new method of network tomography called NTRS (network tomography based on reduction strategy). This method proposes a number of pre-treatment principles to optimize the measurement topology and these principles include merging the consistent physical links, making full use of the known performance information of internal links, narrowing the scope of the performance measuring area and covering the key links. According to the results of measurement, this method screens for link collection with better performance to achieve an effective reduction of the measuring links through computing the covariance of performance parameters for internal links in light of end-to-end performance measurement data. And the method largely improves the accuracy of performance measurement and the link diagnosis speed. Finally, this paper validates the effectiveness of NTRS by simulation, the results show that NTRS can decrease the number of the link performance parameters needed to measure and reduce the computational complexity.

Key words: performance measurement, network tomography, link reduction, system area network

摘要： 系统域网络是高性能计算机、数据中心的重要组成部分，当前系统域网络存在网络规模庞大，内部链路繁多，网络流量行为复杂和各种应用对网络性能状况敏感度高等特点，致使采用传统网络断层扫描方法进行性能测量的计算复杂度呈指数级增长。针对上述问题提出了一种基于消减策略的网络断层扫描方法（network tomography based on reduction strategy，NTRS）。该方法提出了预处理原则对实际物理拓扑进行策略约束，充分利用内部链路已知的性能信息缩小性能测量网络区域和关键链路覆盖，依据测量结果计算链路性能协方差，筛除性能状况较好的链路集合，实现链路数量的有效约简，进而很大程度上提高了网络诊断的准确性。通过模拟实验验证了NTRS方法的有效性，实验结果表明该方法能缩小链路性能参数的测量规模，降低计算的复杂度。

关键词: 性能测量, 网络断层扫描, 链路消减, 系统域网络