pathChirp Spatio-Temporal Available Bandwidth Estimation Vinay Ribeiro Rolf Riedi, Richard Baraniuk Rice University

A Bird’s Eye View of the Internet Data transmitted as packets Multiple routers on end-to-end path Routers queue bursts of packets

Edge-based Probing Internet provides connectivity Lack of optimization Difficult to obtain information from routers Solution: inject probe packets to measure internal properties

Questions awaiting Answers What does the Internet topology look like? Where does congestion occur and why? Given several mirror sites to download data which one to choose? Is my ISP honoring the service-level agreement?

Available Bandwidth Estimation

Network Model Packet delay = constant term (propagation, service time) + variable term (queuing delay) End-to-end paths –Multi-hop –No packet reordering Router queues –FIFO –Constant service rate

Available Bandwidth Unused capacity along path Available bandwidth: Goal: use end-to-end probing to estimate available bandwidth

Applications Network monitoring Server selection Route selection (e.g. BGP) SLA verification Congestion control

Available Bandwidth Probing Tool Requirements Fast estimate within few RTTs Unobtrusive introduce light probing load Accurate No topology information (e.g. link speeds) Robust to multiple congested links No topology information (e.g. link speeds) Robust to multiple congested links

Principle of Self-Induced Congestion Advantages –No topology information required –Robust to multiple bottlenecks TCP-Vegas uses self-induced congestion principle Probing rate < available bw  no delay increase Probing rate > available bw  delay increases

Trains of Packet-Pairs (TOPP) [Melander et al] Vary sender packet-pair spacing Compute avg. receiver packet-pair spacing Constrained regression based estimate Shortcoming: packet-pairs do not capture temporal queuing behavior useful for available bandwidth estimation Packet-pairs Packet train

Pathload [Jain & Dovrolis] CBR packet trains Vary rate of successive trains Converge to available bandwidth Shortcoming Efficiency: only one data rate per train

Chirp Packet Trains Exponentially decrease packet spacing within packet train Wide range of probing rates Efficient: few packets

Chirps vs. Packet-Pairs Each chirp train of N packets contains N-1 packet pairs at different spacings Reduces load by 50% –Chirps: N-1 packet spacings, N packets –Packet-pairs: N-1 packet spacings, 2N-2 packets Captures temporal queuing behavior

Chirps vs. CBR Trains Multiple rates in each chirping train –Allows one estimate per-chirp –Potentially more efficient estimation

CBR Cross-Traffic Scenario Point of onset of increase in queuing delay gives available bandwidth

Bursty Cross-Traffic Scenario Goal: exploit information in queuing delay signature

PathChirp Methodology I.Per-packet pair available bandwidth, (k=packet number) II.Per-chirp available bandwidth III.Smooth per-chirp estimate over sliding time window of size

Self-Induced Congestion Heuristic Definitions: delay of packet k inst rate at packet k

Excursions Must take care while using self-induced congestion principle Segment signature into excursions from x-axis Valid excursions are those consisting of at least “L” packets Apply only to valid excursions

Setting Per-Packet Pair Available Bandwidth Valid excursion increasing queuing delay Valid excursion decreasing queuing delay Last excursion Invalid excursions

pathChirp Tool UDP probe packets No clock synchronization required, only uses relative queuing delay within a chirp duration Computation at receiver Context switching detection User specified average probing rate open source distribution at spin.rice.edu

Performance with Varying Parameters Vary probe size, spread factor Probing load const. Mean squared error (MSE) of estimates Result: MSE decreases with increasing probe size, decreasing spread factor

Multi-hop Experiments First queue is bottleneck Compare –No cross-traffic at queue 2 –With cross-traffic at queue 2 Result: MSE close in both scenarios

Internet Experiments 3 common hops between SLAC  Rice and Chicago  Rice paths Estimates fall in proportion to introduced Poisson traffic

Comparison with TOPP 30% utilization Equal avg. probing rates for pathChirp and TOPP Result: pathChirp outperforms TOPP 70% utilization

Comparison with Pathload 100Mbps links pathChirp uses 10 times fewer bytes for comparable accuracy Available bandwidth EfficiencyAccuracy pathchirppathloadpathChirp 10-90% pathload Avg.min-max 30Mbps0.35MB3.9MB19-29Mbps16-31Mbps 50Mbps0.75MB5.6MB39-48Mbps39-52Mbps 70Mbps0.6MB8.6MB54-63Mbps63-74Mbps

Tight Link Localization

Key Definitions Goal: use end-to-end probing to locate tight link in space and over time Path available bandwidth Sub-path available bandwidth Tight link: link with least available bandwidth

Applications Network monitoring - locating hot spots Network aware applications - server selection Science: where do Internet tight links occur and why?

Methodology Estimate A[1,m] For m>tight link, A[1,m] remains constant

Principle of Self-Induced Congestion Probing rate = R, path available bandwidth = A Advantages –No topology information required –Robust to multiple bottlenecks R < A  no delay increase R > A  delay increases

Packet Tailgating Large packets of size P (TTL=m) small packets of size p Large packets exit at hop m Small packets reach receiver with timing information Previously employed in capacity estimation

Estimating A[1,m] Key: Probing rate decreases by p/(p+P) at link m Assumption: r<A[m+1,N], no delay change after link m R < A[1,m]  no delay increase R > A[1,m]  delay increases

Tight Link Localization Tight link: link after which A[1,m] remains constant Applicable to any self-induced congestion tool: pathload, pathChirp, IGI, netest etc.

ns-2 Simulation Heterogeneous sources Tight link location changes over time pathChirp tracks tight link location change accurately tight link estimate

Internet Experiment Two paths: UIUC  Rice and SLAC  Rice Paths share 4 common links Same tight link estimate for both paths SLAC  Rice tight link UIUC  Rice tight link

Comparison with MRTG Data A[1,m] decreases as expected Tight link location differs from MRTG data by 1 hop SLAC  RiceUIUC  Rice

High Speed Probing System I/O limits probing rate On high speed networks:  cannot estimate A using self-induced congestion

Receiver System I/O Limitation Treat receiver I/O bus as an extra link Use packet tailgating If then we can estimate A[1,N-1]

Sender System I/O Limitations Combine sources to increase net probing rate Issue: machine synchronization

Conclusions Towards spatio-temporal available bandwidth estimation Combine self-induced congestion and packet tailgating Available bandwidth and tight link localization in space and over time ns-2 and Internet experiments encouraging Solutions to system I/O bandwidth limitations spin.rice.edu