Long-Term Forecasting of Internet Backbone Traffic

1
Long-Term Forecasting of Internet Backbone Traffic

• Dina Papagiannaki
• with Nina Taft, Zhi-Li Zhang, Christophe Diot

2
Why is it important?

• Current best practices for IP traffic forecasting
  rely on marketing predictions
• Backbone links large fraction of network
  operators investment
• They have large provisioning cycles (between 6
  and 18 months).
• Current practices can be greatly enhanced using
  historical network measurements

3
Where/When in the backbone?

• Goal Where and When links have to be
  upgraded/added in the core of an IP backbone
  network
• Where?
• Measure traffic aggregate between adjacent PoPs
• When?
• We provide the forecast for current trends
• Operators decide when based on SLAs, current
  provisioning practices, etc.

4
Methodology Roadmap
5
Sprint IP topology
6
Observation 1 Periodicities at 12 and 24 hour
cycle
7
Observation 2 Long-Term Trend and Spikes
8
Methodology Roadmap
9
Wavelet Multiresolution Analysis (MRA)

• Decompose into trend plus details at different
  time scales (time scale as power of 2).
• Finest time scale 90 minutes
• Coarsest time scale 96 hours
• à-trous wavelet transform until 6th timescale
  (261.5 hours96 hours) using B3 spline filter.

10
Wavelet Decomposition
Approximations
Details
11
Methodology Roadmap
12
Reducing the model

• Overall trend accounts for 95-97 of total
  energy
• The maximum amount of energy in the details is
  located at the 3rd timescale (i.e. 12 hours)

13
Analysis of Variance

• Accounts for 80-94 of total variance
• Time series can be easily further compacted into
  weekly time series

14
Methodology Roadmap
15
Forecasting weekly components l(j) and dt3(j)

• Autoregressive Integrated Moving Average models
• Box-Cox methodology for fitting
• Evaluation based on standard fitting indices
• Traffic forecast derived through the model

16
Evaluation of Forecasts
17
Benefits

• Highly accurate forecasts.
• Minimal computational complexity.
• The technique focuses on the aspects of the
  traffic that need to be modeled for the purpose
  of capacity planning.
• The time series analyzed are significantly
  smaller than the initial ones.
• Direct application of Box-Cox methodology leads
  to highly inaccurate forecasts on initial data.

18
Future Work

• Forecasting IP traffic matrices
• As individual OD pairs?
• Or perhaps principal components?
• Are eigenvectors stable across time?
• Issue what do we do about sampling?

19
Questions?
Thank you!