NetQuest: A Flexible Framework for Internet Measurement Lili Qiu Joint work with Mike Dahlin, Harrick Vin, and Yin Zhang UT Austin

2 C&W UUNet Sprint Qwest Earthlink AOL SBC AT&T Server Motivation

3 C&W UUNet AOL Qwest Earthlink Sprint SBC AT&T Motivation (Cont.) Why is it so slow?

4 Motivation (Cont.) Applications are performance-aware –Server selection –Fault diagnosis –Traffic engineering –Overlay networks –Peer-to-peer applications –… Internet: large & decentralized Network measurement is important to –ISPs –Enterprise and university networks –Application and protocol designers –End users –…

5 Key Requirements Scalable: work for large networks (100 –10000 nodes) Flexible: accommodate different applications –Multi-user design Multiple users interested in different parts of network or have different objective functions –Augmented design Conduct additional experiments given existing observations, e.g., after measurement failures –Differentiated design Different quantities have different importance, e.g., a subset of paths belong to a major customer Q: Which measurements to conduct to estimate the quantities of interest?

6 What We Want A function f(x) of link performance x –We use a linear function f(x)=F*x in this talk x1x1 x 11 x4x4 x5x5 x 10 x9x9 x7x7 x6x6 x8x8 3 x2x2 x3x3 Ex. 1: average link delay f(x) = (x1+…+x11)/11 Ex. 2: end-to-end delays Apply to any additive metric, eg. Log (1 – loss rate)

7 Problem Formulation What we can measure: e2e performance Network inference –Given e2e performance, infer link performance –Infer x based on y=F*x, y, and F Design of measurement experiments –State of the art Probe every path (e.g., RON) Rank-based approach [sigcomm04] –Select a “best” subset of paths to probe so that we can accurately infer f(x) –How to quantify goodness of a subset of paths?

8 Bayesian Experimental Design Notations –D: a measurement design (eg., a subset of paths to probe) –I: an inference algorithm –U(D,I): utility function for design D and inference I A good design maximizes the expected utility under the optimal inference algorithm

9 Design Criteria Let, where is covariance matrix of x Bayesian A-optimality –Goal: minimize the squared error Bayesian G*-optimality –Goal: minimize the worst-case squared error Bayesian D-optimal –Goal: maximize the expected gain in Shannon information

10 Flexibility Multi-user design –New design criteria: a linear combination of different users’ design criteria Augmented design –Ensure the newly selected paths in conjunction with previously monitored paths maximize the utility Differentiated design –Give higher weights to the important rows of matrix F

11 Evaluation Methodology Data sets –NLANR traces RTT, loss, traceroute measurements between pairs of 140 universities in Oct –Resilient overlay network (RON) RTT and loss among hosts in March & May 2001 Accuracy metric

12 Evaluation Results (Cont.) Rank-based All pairwise

13 Evaluation Results (Cont.) Rank-based All pairwise

14 Summary of Other Results Bayesian experimental design can support –Multi-user design –Augmented design –Differentiated design Inference accuracy also depends on –Inference algorithms –Prior information

15 Summary Our contributions –Bring Bayesian experimental design to network measurement –Develop a flexible framework to accommodate different design requirements –Experimentally show its effectiveness On-going work –Build a toolkit –Gain operational experience –Develop applications anomaly detection performance knowledge plane

16 Thank you!