Name
Title | ||
|---|---|---|
Scalable diagnosis in IP networks using path-based measurement and inference: A learning framework |
Abstract | ||
|---|---|---|
In this paper, we investigate scalability and performance of measurement-based network monitoring, focusing on failure and congestion diagnosis in IP networks for network-based multimedia applications. Path-based measurements using unicast probe-packets are obtained at end-hosts, and diagnosis is performed by exploiting the spatial dependence among those measurements. We formulate network monitoring in a machine learning framework using probabilistic graphical models which perform inference of the network states (on/off) using unicast measurements. We provide fundamental limits on the relationship between the number of probe packets, the size of a network and the ability to diagnose either failed links or congested network components. Specifically, the diagnosis problem is dealt in a two-fold manner. Initially for fault diagnosis, we construct a graphical model using a Bayesian belief network for path-based measurements. We then provide a lower bound on the average number of probes per edge for link failure diagnosis using variational inference under ''noisy'' probe measurements. Variational inference provides a feasible approximation to address the number of spatially dependent measurements needed for diagnosis in large networks. We then develop an entropy lower (EL) bound by drawing similarities between coding over a binary symmetric channel (BSC) and link failure diagnosis. Both bounds show that the number of measurements needed for diagnosis grows linearly with respect to the number of links. The analytical results are validated by simulation. On the other hand, for congestion diagnosis, we propose a solution based on decoding of linear error control codes on a BSC. In this scenario, we consider path-based probing experiments under both noiseless and ''noisy'' measurements and compare its performance against the fundamental limits. To identify the congested nodes we construct a factor graph, and congestion is inferred using belief-propagation algorithm. Simulation results demonstrate the ability of our approach to perfectly localize congested nodes using a scalable number of measurements and a computationally efficient algorithm. We believe that this study can ease the problem arising due to lack of QoS support and provide good-quality broadband multimedia services. |
Year | DOI | Venue |
|---|---|---|
2010 | 10.1016/j.jvcir.2009.07.007 | J. Visual Communication and Image Representation |
Keywords | Field | DocType |
diagnosis problem,path-based measurement,fault diagnosis,variational inference,link failure diagnosis,congestion diagnosis,scalable diagnosis,ip network,average number,bayesian belief network,fundamental limit,measurements,lower bound,machine learning,factor graph,binary symmetric channel,graphical model,bayesian belief networks,spatial dependence,belief propagation,network monitoring | Factor graph,Computer science,Network packet,Algorithm,Quality of service,Bayesian network,Graphical model,Network monitoring,Unicast,Scalability | Journal |
Volume | Issue | ISSN |
21 | 2 | 1047-3203 |
Citations | PageRank | References |
1 | 0.36 | 30 |
Authors | ||
4 | ||
Name | Order | Citations | PageRank |
|---|---|---|---|
| Rajesh Narasimha | 1 | 30 | 6.18 |
| Souvik Dihidar | 2 | 11 | 2.97 |
| Chuanyi Ji | 3 | 812 | 124.04 |
| Steven W. McLaughlin | 4 | 343 | 41.18 |