pathChirp & STAB Measuring Available Bandwidth and Locating Bottlenecks in Packet Networks Vinay Ribeiro Rolf Riedi, Richard Baraniuk Rice University spin.rice.edu

Rice University | SPIN.rice.edu 2 Packet Networks Data transmitted as packets Routers forward packets until destination Routers buffer packets in queues Link bandwidth = maximum data transmission rate (bits/sec) link

Rice University | SPIN.rice.edu 3 Network Expansion Grown in size and importance Crucial for commerce, government, research, … ARPANET 1969 NSFNET 1993

Rice University | SPIN.rice.edu 4 Study Network Properties Properties –connectivity between routers –bandwidth used on different links –queuing delays –statistical properties of packet arrivals Improve network performance –Network design –Use bandwidth resources efficiently –Reduce delays –Assist network-aware applications

Rice University | SPIN.rice.edu 5 Obtaining Network Information is Hard Different parts of Internet owned by different organizations Information sharing difficult –Commerical interests/trade secrets –Privacy Direct measurement –Router performance affected with too much measurement –Tapping links, extra infrastructure, expensive Sheer volume of information –Cannot measure everything Difficulties faced by network administrator Difficulties faced by network user different organizations

Rice University | SPIN.rice.edu 6 Edge-Based Probing Inject probe packets into network Infer internal properties from packet delay End-to-end packet delay = speed of light propagation + queuing delay probe packets

Rice University | SPIN.rice.edu 7 Probing “Uncertainty Principle” Large volume of probe packets –Accurate inference of network properties –Inefficient use of precious bandwidth resources Small volume of probe packets –Less accurate inference –Efficient use of resources Balance tradeoff in accuracy vs. efficiency

Rice University | SPIN.rice.edu 8 Available Bandwidth Link available bandwidth = unused bandwidth on a link Path available bandwidth = smallest available bandwidth of all links of a path Available bandwidth is time-varying Goal: end-to-end probing to estimate path available bandwidth Link bandwidth = 100Mbps Bandwidth used to transmit packets = 30Mbps Link available bandwidth = 70Mbps 70Mbps 30Mbps 50Mbps 20Mbps 60Mbps Link available bandwidths Example:

Rice University | SPIN.rice.edu 9 Applications Server selection Route selection (e.g. BGP, overlay networks) Service verification Tuning transport protocols UDP-storm attack detection Early warning of meltdown

Rice University | SPIN.rice.edu 10 Probing Tool Requirements Fast, real-time estimate Accurate Efficient, introduce light probing load No topology assumptions (e.g. link bandwidths)

Rice University | SPIN.rice.edu 11 Self-Induced Congestion Advantages –No topology information required Transition point gives estimate of available bandwidth Probing bit rate > available bandwidth  delay increases (queues start filling up) Probing bit rate < available bandwidth  no delay increase (queues do not fill up) time probe packets low probing rate high probing rate

Rice University | SPIN.rice.edu 12 Chirp Packet Trains Exponentially decrease packet spacing within packet train Simultaneously probe at wide range of probing rates Efficient: few packets Example: Chirp of 25 packets with  =1.2 has probing range Mbps (bits/sec)

Rice University | SPIN.rice.edu 13 Available Bandwidth estimation with pathChirp Segment delay profile into increasing/decreasing regions Apply principle of self-induced congestion to each region Average over different regions for per-chirp estimate Final estimate: moving-average of per-chirp estimates

Rice University | SPIN.rice.edu 14 Gigabit Testbed Experiment CAIDA/CalNGI bandwidth estimation lab Vary available bandwidth using cross-traffic generator pathChirp tracks available bandwidth well Mbps time (seconds)

Rice University | SPIN.rice.edu 15 Thin Links Thin link – link with less available bandwidth than all preceding links Sub-path available bandwidth A[1,m] = smallest available bandwidth among first m links Goal: use end-to-end probing to locate thin links in space and track changes in location over time 70Mbps 30Mbps 50Mbps 20Mbps 60Mbps Link available bandwidths

Rice University | SPIN.rice.edu 16 Applications Science: where does congestion occur and why? Network aware application –Route around problem spots in Internet Network monitoring/troubleshooting –Locating hot spots

Rice University | SPIN.rice.edu 17 Estimating Sub-Path Available Bandwidth A[1,m] Replace each packet by two packets: Big packet size P, small packet size p Key: Probing rate decreases by p/(p+P) at link m Self-induced congestion only up to link m Small packets carry timing information to receiver 1 2 m

Rice University | SPIN.rice.edu 18 Tight Link Localization with STAB Thin links: links at which A[1,m] decreases Last thin link has least available bandwidth among all links Implemented in Spatio-Temporal Available Bandwidth estimator (STAB)

Rice University | SPIN.rice.edu 19 Simulation STAB tracks thin links well Actual Estimated Probability that different links are thin links topology t=360 sec t=180 sec Link number m Sub-path available Bandwidth A[1,m] (Mbps) time (sec) Sub-path available Bandwidth A[1,m] (Mbps) Link number m

Rice University | SPIN.rice.edu 20 Probability that different links are thin links Locate thin links on two paths simultaneously Estimated thin link locations are consistent for two paths Internet Experiment time Link number m Sub-path available Bandwidth A[1,m] (Mbps) Sub-path available Bandwidth A[1,m] (Mbps) Router data supports STAB results UIUC  Rice UWisc  Rice

Rice University | SPIN.rice.edu 21 New Research Directions Spatio-temporal network tomography Wireless network probing

Rice University | SPIN.rice.edu 22 Other Projects Synthesis of fractal data Alpha-Beta analysis of Internet data High-speed transport protocols

Rice University | SPIN.rice.edu 23 Synthesis of Fractal Data Bytes/time time series from an Internet link Classical Models (Markov/Poisson) Bytes per 600ms Bytes per 60ms Bytes per 6ms Internet data is fractal --- high variability if we zoom-in or zoom-out Fast synthesis using multifractal wavelet model –Useful for simulations –Code available at dsp.rice.edu People: Matthew Crouse, Rolf Riedi, R. Baraniuk

Rice University | SPIN.rice.edu 24 Alpha-Beta Analysis of Internet Data Connection -- set of all packets with a unique source and destination Few connections (alpha) cause most of the “spikes” Implications for designing simulation topologies, queuing analysis, congestion control People: Shriram Sarvotham, Rolf Riedi, Richard Baraniuk =+ Time series of bytes per 500ms Alpha component “Spiky” Few connections Beta component Gaussian Most connections

Rice University | SPIN.rice.edu 25 High-Speed Transport Protocols Transport protocols – send at maximum data rate that does not congest network Current protocol (TCP-Reno) cannot utilize all the bandwidth on high-speed Giga-bit networks Existing solutions for high-speed networks too aggressive –Negative impact on competing TCP-Reno connections –Cannot deploy such solutions Hybrid protocol –Utilizes bandwidth on high-speed networks –Competes fairly with TCP-Reno connections People: Ryan King, Rolf Riedi, Richard Baraniuk

Rice University | SPIN.rice.edu 26 Conclusions pathChirp – efficient probing tool to estimate path available bandwidth STAB – probing tool to locate thin links in space and track changes in location over time Code (UNIX) – Available for download at spin.rice.edu Other projects – synthesis of fractal data (dsp.rice.edu), alpha-beta analysis, high-speed transport protocols