Accurate Shared Bottleneck Detection Based On SVD and Outliers Detection
msra(2009)
摘要
We present a new mechanism for detecting shared bottlenecks between end-to-end paths in a network. Our mech- anism, which only needs one-way delays from endpoints as an input, is based on the well known linear algebraic approach SVD (Singular Value Decomposition). Clusters of flows which share a bottleneck are extracted from SVD results by applying an outlier detection method. Simulations with varying topologies and different network conditions show the high accuracy of our technique. ND-TO-END flows in packet based networks adversely influence each other when they traverse a network link which is a single point of congestion. We call such a link a "shared bottleneck". Given the large number of distributed applications that are now used on the Internet, where coor- dination between such end-to-end flows is quite feasible, we believe that the detection of shared bottlenecks in the network is one of the key measurement efforts that can be undertaken, as there are numerous possibilities for using such knowledge. Consider, for example, a transfer delay prediction tool (which is a common element of computational Grids) which tells two senders A and B that the time to transfer files to receivers C and D will be four and five minutes, respectively, but fails to inform them that the A-C and B-D path share a bottleneck. Then, both systems would perhaps start to send the files at the same time, and the predicted transfer delay would clearly be wrong because of adverse interactions between the flows at the shared bottleneck. In a Grid, it is not uncommon for a single entity (or person) to be in control of the two senders A and B — thus, with knowledge about shared bottlenecks, this problem might have been prevented by accordingly scheduling the time of these transfers (this is what schedulers in Grid workflow systems do). As another example, one could imagine an overlay network where a file is sent from a single source to a single destinatio n via multiple paths across the overlay in order to speed up the transfer time. Such a scheme would probably only yield a noticeable benefit if the overlay paths that are used do not include a shared bottleneck in the underlay. Similarly,
更多查看译文
关键词
singular value decomposition,network monitor- ing,passive measurement,outliers detection.,index terms—shared bottleneck detection,distributed application,linear algebra,overlay network,network monitoring,indexing terms,outlier detection
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络