single-hop probing asymptotics in available bandwidth...
TRANSCRIPT
![Page 1: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/1.jpg)
1
Single-Hop Probing Asymptotics in Available Bandwidth Estimation: Sample-Path Analysis
SingleSingle--Hop Probing Asymptotics in Hop Probing Asymptotics in Available Bandwidth Estimation: Available Bandwidth Estimation: SampleSample--Path AnalysisPath Analysis
Xiliang Liu, Kaliappa Ravindran, Benyuan LiuCity University of New York
Dmitri LoguinovTexas A&M University
![Page 2: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/2.jpg)
2
OutlineOutlineOutline
• Introduction━ Constant -rate fluid cross-traffic model━ Relationship to existing techniques
• General Bursty Cross-Traffic Model• Experimental Verifications• Implications to Existing Techniques• Conclusion
Prior work
Our work
![Page 3: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/3.jpg)
3
ProblemProblemProblem
• Measuring path avail-bw using probing streams:
• Basic question: the relationship between input, output, and the measurement goal: avail-bw
Internet path
Input cross-traffic Output cross-traffic
Input probing streams Output probing streams
![Page 4: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/4.jpg)
4
Single-Hop Fluid Model 1SingleSingle--Hop Fluid Model 1Hop Fluid Model 1
• Assuming Constant-rate Fluid Cross-traffic━ Constant Cross-traffic intensity λ in any time-interval━ Constant Avail-bw A = C — λ in any time-interval
• Probing rate/gap of packet train━ Probing gap: g━ Probing rate: r = s/g
• Fluid models:
P1Pn P3 P2
(n—1)g
![Page 5: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/5.jpg)
5
Single-Hop Fluid Model 2SingleSingle--Hop Fluid Model 2Hop Fluid Model 2g O
gIs/(C-λ) C-λ rI
r O
![Page 6: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/6.jpg)
6
How Existing Techniques Relate to Fluid ModelsHow Existing Techniques Relate to How Existing Techniques Relate to Fluid ModelsFluid Models
Is this model still valid in general bursty cross-traffic ?- The answer is NOT EXACTLY
PTR
C
TOPP uses the sub linear
segment
Spruce
C-λ rI
rO
![Page 7: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/7.jpg)
7
OutlineOutlineOutline
• Introduction━ Constant -rate fluid cross-traffic model━ Relationship to existing techniques
• General Bursty Cross-Traffic Model• Experimental Verifications• Implications to Exiting Techniques• Conclusion
Prior work
Our work
![Page 8: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/8.jpg)
8
Extending to Bursty Cross-TrafficExtending to Bursty Extending to Bursty CrossCross--TrafficTraffic
• For the gap model, we adapt it to
• gO now varies, we change it to the asymptotic average
• Cross-traffic rate is no longer a constant, λ is interpreted as its long-term average.
![Page 9: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/9.jpg)
9
Real Asymptotic Model Real Asymptotic Model Real Asymptotic Model
• With proof, we offer the following gap model in bursty cross-traffic:
• The two additional terms are zero in fluid traffic, but are often POSITIVE in bursty cross-traffic.
![Page 10: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/10.jpg)
10
What is the term E[Rn(t)]/(n-)What is the term What is the term EE[[RRnn((tt)]/()]/(nn--))Q
ueui
ng d
elay
timet t+gI
R(t)
•E[Rn(t)] is the asymptotic time average of Rn(t)
• Rn(t) is the additional queuing delay imposed on the last packet Pn by the first n- packets in the same probing train when the train arrives into the hop at time t. It is called intrusion residual.
![Page 11: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/11.jpg)
11
What is the termWhat is the termWhat is the term
Asymptotic time average of the hop idle time within the measurement interval of a packet train.
The amount of hop idle time in that measurement interval after the hop is visited by the packet train at time t.
The measurement interval of a packet train when it arrives to the hop at time t.
![Page 12: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/12.jpg)
12
What is the termWhat is the termWhat is the termQ
ueui
ng d
elay
timet t+gI
Measurement interval [t, t+gI]
![Page 13: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/13.jpg)
13
Probing BiasProbing BiasProbing Bias
•The following two terms, called probing bias, are the difference between fluid model and real asymptotic model.
•The closed-form expression of probing bias is given in the paper.
![Page 14: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/14.jpg)
14
Probing Bias VS. Input Gap gIProbing Bias VS. Input Gap gProbing Bias VS. Input Gap gII
gI
Prob
ing
Bias
s/C s/A
In (0, s/C), bias=0
In (s/C, s/A), bias monotonically increasing, but slope <1
When gI>s/A, bias monotonically decreases and asymptotically converges to 0.
![Page 15: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/15.jpg)
15
Gap Model in Bursty Cross-trafficGap Model in Bursty CrossGap Model in Bursty Cross--traffictraffic
gI
E[gO]
Fluid gap model is the lower bound of the real gap model
also identified a upper bound of the gap model
In (0,s/C), real model=fluid model
From s/C beyond, real model starts deviating the fluid model
At s/A, real model reaches the max deviation
After s/A, the real model asymptotically approaches the fluid model
s/C s/A
![Page 16: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/16.jpg)
16
Rate Model in Bursty Cross-trafficRate Model in Bursty CrossRate Model in Bursty Cross--traffictraffic
rI
s/E
[gO]
Fluid rate model becomes an upper bound of the real model
The onset point of deviation is the turning point, smaller then A
Avail-bw point A, where deviation is maximized
Capacity rate C, end of deviation
C-λ C
![Page 17: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/17.jpg)
17
Impact of Packet-train ParametersImpact of PacketImpact of Packet--train Parameterstrain Parameters
• Larger packet size pushes the real model closer to the fluid model━ Sampling interval increases, cross-traffic variance
decreases, cross-traffic is more like fluid.
• Longer packet train also pushes the real model closer to the fluid model.━ Non-intuitive, the paper offers an explanation using
random walk theory.
• Fluid models are tight bounds for real models
![Page 18: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/18.jpg)
18
OutlineOutlineOutline
• Introduction━ Constant -rate fluid cross-traffic model━ Relationship to existing techniques
• General Bursty Cross-Traffic Model• Experimental Verifications• Implications to Exiting Techniques• Conclusion
Prior work
Our work
![Page 19: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/19.jpg)
19
Period Testing 1Period Testing 1Period Testing 1
• The deviation phenomena were first observed in periodic cross-traffic such as CBR
• E[gO] can be easily computed , since it is equal to the time average of gO(t) in one period:
━ Where gO(t) is the output gap of a packet train when it arrives at the hop at time t.
━ Notice that gO(t) is also a periodic function of time with the same period T as that of the cross-traffic.
![Page 20: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/20.jpg)
20
Period Testing 2Period Testing 2Period Testing 2
• Period Testing approximates the time average of gO(t) in [0, T] ━ By sampling it at a set of equally spaced time
instances and taking the average of those samples.
• The number of samples is chosen so that━ Using more samples makes little difference━ Results agree with fluid model when 0<gI<s/C
![Page 21: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/21.jpg)
21
Packet-Pair Rate Curve in CBRPacketPacket--Pair Rate Curve in CBRPair Rate Curve in CBRCBR cross-traffic with average intensity 2.5mb/s, Hop capacity C=10mb/s
As probe packet size increases, both the deviation range and deviation amplitude shrink.
![Page 22: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/22.jpg)
22
Packet-train Rate Curve in CBRPacketPacket--train Rate Curve in CBRtrain Rate Curve in CBRKeep probing packet size to 50bytes, change packet-train length
As packet train length increases, the deviation from fluid upper bound decreases.
![Page 23: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/23.jpg)
23
Trace-driven Testing TraceTrace--driven Testing driven Testing
• Allows examining the asymptotic model in different types of cross-traffic
• Use time average of gO(t) in a finite time interval [0, α] to approximate E[gO]
• α is chosen so that the cross-traffic intensity in [0, α] is close to its long term average
• gO(t) can be computed based on cross-traffic trace and hop capacity C, when t+(n-1)gI< α
![Page 24: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/24.jpg)
24
Cross-traffic TracesCrossCross--traffic Tracestraffic TracesA
vera
ge In
tens
ity in
[0, t
] (m
b/s)
t (s)
All 4 traces have long-term intensity 3mb/s Use 20s trace for Poisson traffic and 60s trace for Pareto on/off traffic
![Page 25: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/25.jpg)
25
Packet-Pair Rate CurvesPacketPacket--Pair Rate CurvesPair Rate Curves
•Probing packet size=750bytes.
•CBR rate curve overlap with Pareto on/off rate curve
•PUS (yellow) suffers slightly more bias then PCS
![Page 26: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/26.jpg)
26
Packet-Train Rate CurvesPacketPacket--Train Rate CurvesTrain Rate Curves
•Use 16-packet probing train, packet size=750bytes.
•CBR curve becomes almost unbiased.
•The other three overlap.
![Page 27: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/27.jpg)
27
Probing bias VS. Cross-traffic BurstinessProbing bias VS. CrossProbing bias VS. Cross--traffic Burstinesstraffic Burstiness
• The results so far shows that:━ As probing packet size or train length increases,
probing bias vanishes.━ The vanishing rate depends on cross-traffic
burstiness. CBR>Poisson>Pareto on/off━ Although Pareto on/off is more bursty than
Poisson, at certain time interval, the traffic variance can be smaller than Poisson, causing less probing bias in its rate curve.
• More discussion is in the paper
![Page 28: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/28.jpg)
28
OutlineOutlineOutline
• Introduction━ Constant -rate fluid cross-traffic model━ Relationship to existing techniques
• General Bursty Cross-Traffic Model• Experimental Verifications• Implications to Exiting Techniques• Conclusion
Prior work
Our work
![Page 29: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/29.jpg)
29
Implication to existing techniques 1Implication to existing techniques 1Implication to existing techniques 1
• TOPP use a transformed rate curve which is piece-wise linear in fluid cross-traffic
• Real asymptotic curves are not the same as the fluid models. This can cause significant under estimation of avail-bw even in a single-hop path
![Page 30: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/30.jpg)
30
Implication to existing techniques 2Implication to existing techniques 2Implication to existing techniques 2
Applying TOPP’s linear regression on the real curves will produce avail-bw estimation:
5.023.283.436.67
POFPUSPCSCBR
![Page 31: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/31.jpg)
31
Implication to existing techniques 3Implication to existing techniques 3Implication to existing techniques 3
• Searching for the turning point (PTR) as available bandwidth causes negative bias━ However, this bias can be mitigated to negligible
level using long packet train.
• Sampling cross-traffic (Spruce) with rI≥C is unbiased in single-hop path━ At this input rate, the real model agrees with fluid
model.
![Page 32: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/32.jpg)
32
ConclusionsConclusionsConclusions
• We developed an understanding of single-hop bandwidth estimation in busty cross-traffic that extends prior fluid models
• Cross-traffic burstiness implies bandwidth underestimation to several existing techniques. The underestimation can be mitigated using long train and large packet size
• Future work is to extend our understanding to multi-hop bandwidth estimation
![Page 33: Single-Hop Probing Asymptotics in Available Bandwidth ...irl.cs.tamu.edu/people/xiliang/papers/imc2004-ppt.pdf19 Period Testing 1Period Testing 1 • The deviation phenomena were first](https://reader034.vdocument.in/reader034/viewer/2022051904/5ff59e63021bdd0d8a2bc93c/html5/thumbnails/33.jpg)
33