A Machine Learning Approach to TCP Throughput Prediction

1
A Machine Learning Approach to TCP Throughput
Prediction

• Mariyam Mirza
• Joel Sommers
• Paul Barford
• Xiaojin Zhu

2
Motivation Why Predict TCP Throughput?

• Multiple Paths b/w senders and receivers
• Select best path
• Common definition of best path highest
  throughput path

3
Talk Outline

• Goals and Challenges of TCP Throughput Prediction
• Previous Work
• Our Approach
• Results
• Summary

4
TCP Throughput Prediction Challenges and
Existing Approaches

• Goals or Challenges
• Accuracy
• Timeliness, i.e., Responsiveness to changing
  conditions
• Cost Volume of probe traffic introduced
• Previous approaches
• Formula-Based e.g., Padhye et. al., 1998
• History-Based He et. al., 2005

5
Overview of Our Approach

• Measure path characteristics using lightweight
  probes
• Use Support Vector Regression (SVR), for
  prediction
• Advantages over existing approaches
• Formula-Based (FB)
• Need different formulae for different flavors of
  TCP
• History-Based (HB)
• Heavyweight
• So far, shown to work only for bulk transfers
• Our Probe- and SVR-Based Approach (SVR)
• Lightweight 10x less traffic than HB
• Per-path, so different flavors of TCP
  accommodated automatically, unlike FB
• Wide range of background traffic rates
• Level shifts
• Large and small TCP transfers

6
Experimental Setup
Traffic Generator Hosts
Traffic Generator Hosts
Adtech SX-14 25 ms one-way delay
Cisco 12000
GE
OC-12
GE
GE
OC-3
Cisco 12000
Cisco 12000
Cisco 6500
Cisco 6500
GE
OC-3
GE
Cisco 12000
OC-12
GE
GE
Probe Sender
Endace DAG Monitor with 3.5/3.8 cards, pkts
copied via optical splitters
Probe Sender
7
Path Characteristics Considered

• Highly Accurate Oracular Passive Measurements
• Practical Active Measurements
• Loss (L)
• Loss Frequency
• Loss Duration
• Active Measurements via Badabing Sommers et.
  al., 2005
• Queuing Delay (Q)
• Active Measurements via Badabing
• Available Bandwidth (AB)
• Active Measurements via Yaz Sommers et. al.,
  2006

8
Support Vector Regression (SVR)

• State-of-the-art machine learning tool for
  multivariate regression
• Input features to the SVR AB, Q, L
• Use a Radial Basis Function (RBF) Kernel
• Training produces a highly non-linear prediction
  function
• Non-linearity captures the complex relationship
  between throughput and measurements AB, Q, L
• Apply function to test input features to get
  predictions

9
Experimental Protocol
Yaz run for one AB measurement, 10-30 sec to
converge
File Transferred Oracular measurements of Q, L,
and AB
Badabing run for 30 sec, Q L measured
time
One Experiment

• Training Set and Test Set
• 100 Experiments per set
• Background Traffic 135Mbps, generated by Harpoon
  Sommers et. al., 2004
• Bottleneck Bandwidth OC-3, 150Mbps

10
Results HB Prediction (Baseline)
Predicted Throughput, Mbps
Actual Throughput, Mbps
11
AB-based Predictions,Oracular Passive
Measurements
Predicted Throughput, Mbps
Actual Throughput, Mbps
12
Q-based Predictions, Oracular Passive Measurements
Predicted Throughput, Mbps
Actual Throughput, Mbps
13
L-Based Predictions, Oracular Passive
Measurements
Predicted Throughput, Mbps
Actual Throughput, Mbps
14
Best Prediction Results Q- and L-Based, Oracular
Predicted Throughput, Mbps
Actual Throughput, Mbps
15
Best ResultsQ- and L-Based, Oracular and
Practical
Predicted Throughput, Mbps
Actual Throughput, Mbps
16
Results Summary

• Available Bandwidth not necessary for accurate
  throughput prediction
• Relative Error He et. al., 2005
• Relative Error Predicted Throughput - Actual
  Throughput
• min(predicted throughput, actual throughput)

17
Results Level Shifts, Practical Measurements
18
Results Different File Sizes
Predicted Throughput, Mbps
Training Sizes 32KB, 512KB, 8MB Test Sizes
Randomly Generated, b/w 2KB and 8MB
Actual Throughput, Mbps
19
PathPerf Online Tool for TCP Throughput
Prediction
Make Prediction Retrain if necessary
Badabing run for 30 sec, Q L measured
File Transferred
time
One Experiment

• Training Set First Measurement to start out with
• Test Set Each new measurement
• Retrain if prediction error exceeds threshold
• Practical Passive Measurements only no Oracular
  Passive Measurements
• No AB measurements
• Will be released soon

20
PathPerf Wide-Area Experiments

• Run on the RON testbed in Dec 2006
• Paths cross-section of 7 nodes
• Nodes in Amsterdam, London, Utah, NYC, Ithaca,
  New Mexico, Maryland
• Base RTT range 10ms-150ms
• Throughput on most paths window limited

21
Sample Wide-Area Result Without Retraining
22
Sample Wide-Area Result With Retraining
23
Summary

• Active Probing and Machine Learning Based
  Mechanism for TCP Throughput Prediction
• Accurate
• Lightweight
• Responsive

24
Acknowledgements

• David Anderson, for help using the RON testbed

25
Questions?