Post on 20-Jan-2016
description
Estimating Link Capacity Estimating Link Capacity in High Speed Networksin High Speed Networks
Ling-Jyh ChenLing-Jyh Chen11, Tony Sun, Tony Sun22, Li Lao, Li Lao22, Guang Yang, Guang Yang22, M.Y. Sanadidi, M.Y. Sanadidi22, Mario Gerla, Mario Gerla22
11Institute of Information Science, Academia SinicaInstitute of Information Science, Academia Sinica22Dept. of Computer Science, University of California at Los AngelesDept. of Computer Science, University of California at Los Angeles
DefinitionDefinition CapacityCapacity: maximum IP-layer throughput that a flow can
get, without any cross traffic. Available BandwidthAvailable Bandwidth: maximum IP-layer throughput that
a flow can get, given (stationary) cross traffic.
Previous Work on Capacity EstimationPrevious Work on Capacity Estimation
Per-hop based pathchar: use different packet sizes to probe the per-
hop link capacity clink, pchar: variants of pathchar Nettimer: use “packet tailgating” technique
End-to-end based Pathrate, Sprobe, CapProbe
For specialized networks: AsymProbe, ALBP, AdHoc Probe
How about high speedhigh speed networks?
Estimating High Speed LinksEstimating High Speed Links
High speed links are becoming popular (e.g. GB links, DVB links, and UWB links)
However, capacity estimation on high speed links remains a challenge (e.g., probing pksize and system time resolution are limited)
And, an effective estimation tool for high speed links is still desired
Our ContributionOur Contribution
We propose an end-to-end capacity estimation technique for high speed links, called PBProbePBProbe.
PBProbe is based on CapProbe One-way method UDP based packet bulkpacket bulk based simple, fast, and accurate
Packet Pair DispersionPacket Pair DispersionT3
T2 T3
T3
T1
T3
Narrowest Link
20Mbps 10Mbps 5Mbps 10Mbps 20Mbps 8Mbps
Capacity = (Packet Size) / (Dispersion)
Issues: Compression and ExpansionIssues: Compression and Expansion
• Queueing delay on the first packet => compression
• Queueing delay on the second packet => expansion
CapProbe CapProbe (Rohit et al, SIGCOMM’04)(Rohit et al, SIGCOMM’04)
Key insight: a packet pair that gets through with zero queueing delay yields the exact estimate.
CapProbe uses “Minimum Delay Sum” filter.
CapacityCapacity
Proposed Approach: PBProbeProposed Approach: PBProbe Have two phases for both forward and
backward link estimation Use packet bulk (instead of packet pair) of
length k in each probing Adapt k to enlarge the dispersion between
the first and last packet, and thus overcome the timer resolution problem
Tradeoff BW consumption and estimation speed by U parameter
Proposed Approach: PBProbeProposed Approach: PBProbe
Proposed Approach: PBProbeProposed Approach: PBProbe k is depended on
the estimated link capacity and the supported timer resolution.
n is set to 200. Dthresh is set to 1ms. U is set to 0.002.
AnalysisAnalysis
Poisson cross traffic (arrival and service rates are λ and μ), service time is τ
Prob. of obtaining a good sample:
Expected number of samples required for obtaining a good sample:
AnalysisAnalysis
EvaluationEvaluation
NISTNet emulation High speed Internet experiments Comparison of PBProbe and Pathrate
Evaluation 1: NISTNet emulationEvaluation 1: NISTNet emulation
No cross traffic
Evaluation 2: Internet experimentsEvaluation 2: Internet experiments 5 hosts: NTNU, UCLA,
CalTech, GaTech, PSC(n = 200, k = 100, 20 runs)
Evaluation 3: PBProbe vs PathrateEvaluation 3: PBProbe vs Pathrate
SummarySummary
We propose an end-to-end capacity estimation technique, called PBProbe, for high speed links.
We evaluate PBProbe by analysis, emulation and Internet experiments.
We show that PBProbe can correctly and rapidly estimate bottleneck capacity in almost all test cases.
AcknowledgementsAcknowledgements
This work is co-sponsored by the National Science Council and the National Science Foundation under grant numbers NSC-94-2218-E-001-002 and CNS-0435515.
We are grateful to the following people for their help in carrying out PBProbe measurements: Sanjay Hegde (CalTech), Che-Chih Liu (NTNU), Cesar A. C. Marcondes (UCLA), and Anders Persson (UCLA).
Thanks!Thanks!
CapProbe:CapProbe: http://nrl.cs.ucla.edu/CapProbe/