decentralized prediction of end-to-end network...
TRANSCRIPT
![Page 1: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/1.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Prediction of End-to-End NetworkPerformance Classes
Yongjun Liao, Wei Du, Pierre Geurts, Guy Leduc
Research Unit in Networking(RUN), University of Liege, Belgium
December 08, 2011
1 / 27
![Page 2: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/2.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
End-to-End Network Performance
Metrics
round-trip time (RTT)
available bandwidth (ABW)
packet loss rate (PLR)
Network Performance Matters!
peer-to-peer downloading
overlay routing
content distribution network
Internet games
Peer Selection
Internet
smallest RTT
highest ABW
2 / 27
![Page 3: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/3.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Acquisition on Large-Scale Networks
full-mesh active measurements
1
23
4
5
67
8
9
n nodes⇒ o(n2) measurements
accurate but expensive
network performance prediction
1
23
4
5
67
8
9
n nodes⇒ measurements� o(n2)
less accurate but cheap
3 / 27
![Page 4: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/4.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Network Performance Prediction
Challenges
Networks are dynamic.I Churn: nodes join and leave frequently.I Metric values vary over time.
Metrics differ largely.I RTT: symmetric; ABW: asymmetric.I RTT: the smaller the better; ABW: the larger the better.I RTT and ABW are measured with different methodologies.
Decentralized processing is prefered.I no landmarks or central serversI no infrastructure
4 / 27
![Page 5: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/5.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Network Performance Prediction
Challenges
Networks are dynamic.I Churn: nodes join and leave frequently.I Metric values vary over time.
Metrics differ largely.I RTT: symmetric; ABW: asymmetric.I RTT: the smaller the better; ABW: the larger the better.I RTT and ABW are measured with different methodologies.
Decentralized processing is prefered.I no landmarks or central serversI no infrastructure
4 / 27
![Page 6: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/6.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Network Performance Prediction
Challenges
Networks are dynamic.I Churn: nodes join and leave frequently.I Metric values vary over time.
Metrics differ largely.I RTT: symmetric; ABW: asymmetric.I RTT: the smaller the better; ABW: the larger the better.I RTT and ABW are measured with different methodologies.
Decentralized processing is prefered.I no landmarks or central serversI no infrastructure
4 / 27
![Page 7: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/7.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Network Performance Prediction
Challenges
Networks are dynamic.I Churn: nodes join and leave frequently.I Metric values vary over time.
Metrics differ largely.I RTT: symmetric; ABW: asymmetric.I RTT: the smaller the better; ABW: the larger the better.I RTT and ABW are measured with different methodologies.
Decentralized processing is prefered.I no landmarks or central serversI no infrastructure
4 / 27
![Page 8: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/8.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Related Work on Network Performance Prediction
Round-Trip Time
Euclidean EmbeddingI GNP: Ng et al. INFOCOM 2002I Vivaldi: Dabek et al. SIGCOMM 2004
Matrix FactorizationI IDES: Mao et al. IMC 2004I DMF: Liao et al. Networking 2010
Available Bandwidth
SEQUOIA: Rama et al. SIGMETRICS 2009
iPlane: Madhyastha et al. USENIX OSDI 2006
5 / 27
![Page 9: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/9.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Our Contributions
1. Class-based Performance Representation
Represent network performance by discrete-valued classes, insteadof real-valued quantities.
2. Formulation as Matrix Completion
Treat the prediction problem as a matrix completion problem.
3. Decentralized Prediction Algorithm
DMFSGD: a decentralized matrix facotrization algorithm basedon stochastic gradient descent.
6 / 27
![Page 10: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/10.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Our Contributions
1. Class-based Performance Representation
Represent network performance by discrete-valued classes, insteadof real-valued quantities.
2. Formulation as Matrix Completion
Treat the prediction problem as a matrix completion problem.
3. Decentralized Prediction Algorithm
DMFSGD: a decentralized matrix facotrization algorithm basedon stochastic gradient descent.
6 / 27
![Page 11: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/11.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Our Contributions
1. Class-based Performance Representation
Represent network performance by discrete-valued classes, insteadof real-valued quantities.
2. Formulation as Matrix Completion
Treat the prediction problem as a matrix completion problem.
3. Decentralized Prediction Algorithm
DMFSGD: a decentralized matrix facotrization algorithm basedon stochastic gradient descent.
6 / 27
![Page 12: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/12.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Our Contributions
1. Class-based Performance Representation
Represent network performance by discrete-valued classes, insteadof real-valued quantities.
2. Formulation as Matrix Completion
Treat the prediction problem as a matrix completion problem.
3. Decentralized Prediction Algorithm
DMFSGD: a decentralized matrix facotrization algorithm basedon stochastic gradient descent.
6 / 27
![Page 13: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/13.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Class-based Performance Representation
Binary Classification
“good” or “bad”
Class reflects the QoS experience of end users.
Class unifies different metrics.
Class information is often sufficient.I Streaming applications care if ABW is high enough.I P2P applications care if RTT is small enough.
Class measurements are cheap.I Classes are rough.I Classes are more stable.
7 / 27
![Page 14: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/14.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Class-based Performance Representation
Binary Classification
“good” or “bad”
Class reflects the QoS experience of end users.
Class unifies different metrics.
Class information is often sufficient.I Streaming applications care if ABW is high enough.I P2P applications care if RTT is small enough.
Class measurements are cheap.I Classes are rough.I Classes are more stable.
7 / 27
![Page 15: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/15.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Class-based Performance Representation
Binary Classification
“good” or “bad”
Class reflects the QoS experience of end users.
Class unifies different metrics.
Class information is often sufficient.I Streaming applications care if ABW is high enough.I P2P applications care if RTT is small enough.
Class measurements are cheap.I Classes are rough.I Classes are more stable.
7 / 27
![Page 16: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/16.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Class-based Performance Representation
Binary Classification
“good” or “bad”
Class reflects the QoS experience of end users.
Class unifies different metrics.
Class information is often sufficient.I Streaming applications care if ABW is high enough.I P2P applications care if RTT is small enough.
Class measurements are cheap.I Classes are rough.I Classes are more stable.
7 / 27
![Page 17: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/17.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Class-based Performance Representation
Binary Classification
“good” or “bad”
Class reflects the QoS experience of end users.
Class unifies different metrics.
Class information is often sufficient.I Streaming applications care if ABW is high enough.I P2P applications care if RTT is small enough.
Class measurements are cheap.I Classes are rough.I Classes are more stable.
7 / 27
![Page 18: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/18.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Measure Performance Classes
“good” or “bad”
Thresholding
Measure if the metric value is larger or smaller than a threshold τ .
If RTT < 100ms, performance is “good”.
If ABW > 100Mbps, performance is “good”.
Measuring classes is much cheaper!
Threshold τ
defined according to requirements of applications
Google TV requires 10Mbps for HD contents.
8 / 27
![Page 19: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/19.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Measure Performance Classes
“good” or “bad”
Thresholding
Measure if the metric value is larger or smaller than a threshold τ .
If RTT < 100ms, performance is “good”.
If ABW > 100Mbps, performance is “good”.
Measuring classes is much cheaper!
Threshold τ
defined according to requirements of applications
Google TV requires 10Mbps for HD contents.
8 / 27
![Page 20: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/20.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Measure Performance Classes
“good” or “bad”
Thresholding
Measure if the metric value is larger or smaller than a threshold τ .
If RTT < 100ms, performance is “good”.
If ABW > 100Mbps, performance is “good”.
Measuring classes is much cheaper!
Threshold τ
defined according to requirements of applications
Google TV requires 10Mbps for HD contents.
8 / 27
![Page 21: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/21.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Our Contributions
1. Class-based Performance Representation
Represent network performance by discrete-valued classes, insteadof real-valued quantities.
2. Formulation as Matrix Completion
Treat the prediction problem as a matrix completion problem.
3. Decentralized Prediction Algorithm
DMFSGD: a decentralized matrix facotrization algorithm basedon stochastic gradient descent.
9 / 27
![Page 22: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/22.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
Matrix Completion
Predict the unknown entries from a few known entries.
Why is it possible?
Matrix entries are correlated.
The correlations induce low rank.
n × n matrix of rank r < nI only r linearly independent
columns or rows
You don’t need all n × n entries!
X
10 / 27
![Page 23: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/23.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
Matrix Completion
Predict the unknown entries from a few known entries.
Why is it possible?
Matrix entries are correlated.
The correlations induce low rank.
n × n matrix of rank r < nI only r linearly independent
columns or rows
You don’t need all n × n entries!
X
10 / 27
![Page 24: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/24.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
Matrix Completion
Predict the unknown entries from a few known entries.
Why is it possible?
Matrix entries are correlated.
The correlations induce low rank.
n × n matrix of rank r < nI only r linearly independent
columns or rows
You don’t need all n × n entries!
X
10 / 27
![Page 25: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/25.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
Matrix Completion
Predict the unknown entries from a few known entries.
Why is it possible?
Matrix entries are correlated.
The correlations induce low rank.
n × n matrix of rank r < nI only r linearly independent
columns or rows
You don’t need all n × n entries!
X
10 / 27
![Page 26: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/26.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
Matrix Completion
Predict the unknown entries from a few known entries.
Why is it possible?
Matrix entries are correlated.
The correlations induce low rank.
n × n matrix of rank r < nI only r linearly independent
columns or rows
You don’t need all n × n entries!
X
10 / 27
![Page 27: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/27.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
11 / 27
![Page 28: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/28.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
3
1 2
6
8
11 / 27
![Page 29: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/29.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
3
1 2
6
8
11 / 27
![Page 30: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/30.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
3
1 2
6
8
good or bad
11 / 27
![Page 31: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/31.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
3
1 2
6
8
1 or -1
11 / 27
![Page 32: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/32.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
1 2
6
8
5
2
8
4
7
11 / 27
![Page 33: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/33.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
1 2
6
8
5
2
8
4
7
11 / 27
![Page 34: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/34.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
1 2
6
8
2
8
4
7
11 / 27
![Page 35: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/35.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Matrix Completion for Network Performance Prediction
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
3
1 2
45
6 7
8
1 2
6
8
2
8
4
7
11 / 27
![Page 36: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/36.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Why is Matrix Completion Possible
Correlations across network performance
network topology
routing algorithms
redundancies among network paths
...
Interneti1
i2
j
12 / 27
![Page 37: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/37.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Why is Matrix Completion Possible
Low-rank of performance matrices
1 5 10 15 200
0.2
0.4
0.6
0.8
1
# singular value
sin
gu
lar
va
lue
s
RTT
RTT class
ABW
ABW class
RTT matrix: 2255× 2255
ABW matrix: 201× 201
Class matrices are obtained by thresholding.
13 / 27
![Page 38: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/38.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Low-Rank Matrix Factorization
≈
X X
Rank(X ) = r
14 / 27
![Page 39: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/39.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Low-Rank Matrix Factorization
≈
X X
Rank(X ) = r
= ×
U V T
14 / 27
![Page 40: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/40.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Low-Rank Matrix Factorization
≈
X X
Rank(X ) = r
= ×
U V T
Look for (U ,V ), instead of X
14 / 27
![Page 41: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/41.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Our Contributions
1. Class-based Performance Representation
Represent network performance by discrete-valued classes, insteadof real-valued quantities.
2. Formulation as Matrix Completion
Treat the prediction problem as a matrix completion problem.
3. Decentralized Prediction Algorithm
DMFSGD: a decentralized matrix facotrization algorithm basedon stochastic gradient descent.
15 / 27
![Page 42: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/42.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Stochastic Gradient Descent
≈
X X
Rank(X ) = r
= ×
U V T
16 / 27
![Page 43: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/43.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Stochastic Gradient Descent
≈
X X
Rank(X ) = r
= ×
U V T
xij xij
ui
vTj
≈ = uivTj
16 / 27
![Page 44: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/44.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
17 / 27
![Page 45: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/45.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
17 / 27
![Page 46: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/46.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
17 / 27
![Page 47: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/47.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
17 / 27
![Page 48: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/48.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
j
17 / 27
![Page 49: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/49.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
j
uj , vj
ui , vi
17 / 27
![Page 50: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/50.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
j
uj , vj
ui , vi
probe(ui )
17 / 27
![Page 51: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/51.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
j
uj , vj
ui , vi
compute xij
17 / 27
![Page 52: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/52.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
j
uj , vj
ui , vi
reply(vj , xij )
17 / 27
![Page 53: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/53.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
j
uj , vj
ui , vi
update vj
uivTj ≈ xij
update ui
uivTj ≈ xij
17 / 27
![Page 54: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/54.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
ui , vi use phase
k
uk , vk
17 / 27
![Page 55: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/55.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
No construction of matrices
X : measurement xij is probed by node i .
U, V : row vectors ui , vi are stored at node i .
Repeated SGD updates
When xij is available, update so that xij ≈ uivTj .
i
ui , vi use phase
k
ui ,vi uk ,vk
17 / 27
![Page 56: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/56.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Decentralized Matrix Factorization by Stochastic Gradient Descent
DMFSGD
All nodes employ the same processing.
Each node selects k neighbors to communicate with.
Each node collaborates with one neighbor at each time.
Advantages
easy to implement
computationally lightweight
suitable for large-scale dynamic measurements
adaptable for various metrics
18 / 27
![Page 57: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/57.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Experiments and Evaluations
Datasets
Harvard Meridian HP-S3
nodes 226 2500 231
metric RTT RTT ABW
dynamic Yes No No
source Ledlie et al. Wong et al. Ramasubramanian et al.NSDI 2007 SIGCOMM 2005 SIGMETRICS 2009
*Harvard dataset is collected from Azureus (now Vuze) and contains dynamicmeasurements with time-stamps.
Accuracy= # of correct prediction# of data
Harvard Meridian HP-S3
Accuracy 89.4% 85.4% 87.3%
19 / 27
![Page 58: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/58.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Experiments and Evaluations
Datasets
Harvard Meridian HP-S3
nodes 226 2500 231
metric RTT RTT ABW
dynamic Yes No No
source Ledlie et al. Wong et al. Ramasubramanian et al.NSDI 2007 SIGCOMM 2005 SIGMETRICS 2009
*Harvard dataset is collected from Azureus (now Vuze) and contains dynamicmeasurements with time-stamps.
Accuracy= # of correct prediction# of data
Harvard Meridian HP-S3
Accuracy 89.4% 85.4% 87.3%
19 / 27
![Page 59: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/59.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Find a “good” peer, not the “best” peer!
Internet
20 / 27
![Page 60: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/60.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Some peers are predicted as “good” and some “bad”.
Internet
,
,
,
/
/20 / 27
![Page 61: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/61.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Select a peer that is predicted as “good”.
Internet
,
,
,
/
/20 / 27
![Page 62: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/62.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
The node is satisfied if the selected peer is truly “good”.
Internet
,
,
,
/
/
satisfied truly “good”
20 / 27
![Page 63: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/63.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
The node is unsatisfied if the selected peer is actually “bad”.
Internet
,
,
,
/
/
unsatisfied actually “bad”
20 / 27
![Page 64: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/64.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Evaluation
Count the number of unsatisfied nodes.
Methods that are compared with
classification: class-based prediction
random peer selection
regression: value-based predictionI Predict values of some metric by our DMFSGD algorithm.I Select the predicted best peer for each node.I Check if the selected peers are truly “good”.
classification with noiseI Overall 15% erroneous labels were simulated.
21 / 27
![Page 65: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/65.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Evaluation
Count the number of unsatisfied nodes.
Methods that are compared with
classification: class-based prediction
random peer selection
regression: value-based predictionI Predict values of some metric by our DMFSGD algorithm.I Select the predicted best peer for each node.I Check if the selected peers are truly “good”.
classification with noiseI Overall 15% erroneous labels were simulated.
21 / 27
![Page 66: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/66.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Evaluation
Count the number of unsatisfied nodes.
Methods that are compared with
classification: class-based prediction
random peer selection
regression: value-based predictionI Predict values of some metric by our DMFSGD algorithm.I Select the predicted best peer for each node.I Check if the selected peers are truly “good”.
classification with noiseI Overall 15% erroneous labels were simulated.
21 / 27
![Page 67: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/67.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Evaluation
Count the number of unsatisfied nodes.
Methods that are compared with
classification: class-based prediction
random peer selection
regression: value-based predictionI Predict values of some metric by our DMFSGD algorithm.I Select the predicted best peer for each node.I Check if the selected peers are truly “good”.
classification with noiseI Overall 15% erroneous labels were simulated.
21 / 27
![Page 68: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/68.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Evaluation
Count the number of unsatisfied nodes.
Methods that are compared with
classification: class-based prediction
random peer selection
regression: value-based predictionI Predict values of some metric by our DMFSGD algorithm.I Select the predicted best peer for each node.I Check if the selected peers are truly “good”.
classification with noiseI Overall 15% erroneous labels were simulated.
21 / 27
![Page 69: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/69.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
Harvard
10 20 30 40 50 600
0.1
0.2
0.3
0.4
0.5
peer number
un
sa
tisfie
d n
od
e p
erc
en
tag
e
Random
Classification
Regression
Classification with noise
22 / 27
![Page 70: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/70.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Peer Selection
HP-S3
10 20 30 40 50 600.05
0.1
0.15
0.2
0.25
0.3
peer number
un
sa
tisfie
d n
od
e p
erc
en
tag
e
Random
Classification
Regression
Classification with noise
22 / 27
![Page 71: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/71.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Conclusions and Future Work
Decentralized Prediction of Network Performance Classes
Class-based Performance Representation
Formulation as Matrix Completion
DMFSGD: a decentralized matrix factorization algorithmby stochastic gradient descent
I accurate and scalableI generic to deal with various metricsI robust against erroneous measurementsI usable on real Internet applications
Future Work
Multiclass classification
, , , ,
23 / 27
![Page 72: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/72.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Acknowledgement
We wish to thank
Dr. Ramasubramanian for providing the HP-S3 dataset;
our shepherd Augustin Chaintreau;
the anonymous reviewers;
project FP7-Fire ECODE.
Thank you for listening. Any questions?
24 / 27
![Page 73: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/73.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Low-Rank Matrix Factorization
(U,V ) = arg min L(X ,U,V ,W , λ)
= arg minn∑
i ,j=1
wij l(xij , uivTj ) + λ
n∑i=1
uiuTi + λ
n∑i=1
vivTi
(U,V ) = {(ui , vi ), i = 1, . . . , n}wij = 1 if xij is known and 0 otherwise
l : loss function that penalizes the difference between x and xI square loss function: l(x , x) = (x − x)2;I hinge loss function: l(x , x) = max(0, 1− xx);I logistic loss function: l(x , x) = ln(1 + e−xx).
λ: regularization coefficient
25 / 27
![Page 74: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/74.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Parameter Sensitivity
parameter tested valueslearning rate η 0.001, 0.01, 0.1, 1
regularization coefficient λ 0.001, 0.01, 0.1, 1rank r 3, 10, 20, 100
loss function l hinge, logisticneighbor number k 5, 10, 30, 50 (Harvard and HP-S3)
16, 32, 64, 128 (Meridian)classification threshold τ 10%, 25%, 50%, 75%, 90%
(portion of good-performing)*chosen value
Insensitive because the inputs are either 1 or −1.
26 / 27
![Page 75: Decentralized Prediction of End-to-End Network …conferences.sigcomm.org/.../Liao-DecentralizedPrediction.pdfDecentralized Prediction of End-to-End Network Performance Classes Yongjun](https://reader035.vdocument.in/reader035/viewer/2022062602/5f0207127e708231d402377e/html5/thumbnails/75.jpg)
Introduction Class-based Representation Matrix Completion Experiments and Evaluations Conclusions and Future Work
Robustness Against Erroneous Measurement
Source of Erroneous Measurements
inaccurate measurement techniques
network anomaly
Errors Type
1 Flip near τ
2 Underestimation bias
3 Flip randomly
4 Good-to-Bad
27 / 27