Transcript Intel Labs Presentation Template

Long-Term Forecasting of Internet
Backbone Traffic
Dina Papagiannaki
with Nina Taft, Zhi-Li Zhang,
Christophe Diot
www.intel.com/research
• Intel Research •
Why is it important?
 Current best practices for IP traffic forecasting rely
on marketing predictions
 Backbone links large fraction of network operator’s
investment
 They have large provisioning cycles (between 6 and
18 months).
 Current practices can be greatly enhanced using
historical network measurements
2
www.intel.com/research
• Intel Research •
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.
3
www.intel.com/research
• Intel Research •
Methodology Roadmap
SNMP
Topology
Traffic
Aggregates
1 signal
Wavelet
Multiresolution
Analysis
7 signals
Model
Reduction
(ANOVA)
2 signals
PoP pair forecast
Individual
Forecasts
(ARIMA)
Weekly time series
4
www.intel.com/research
• Intel Research •
Sprint IP topology
5
www.intel.com/research
• Intel Research •
Observation 1: Periodicities at
12 and 24 hour cycle
6
www.intel.com/research
• Intel Research •
Observation 2: Long-Term
Trend and Spikes
7
www.intel.com/research
• Intel Research •
Methodology Roadmap
SNMP
Topology
Traffic
Aggregates
1 signal
Wavelet
Multiresolution
Analysis
7 signals
Model
Reduction
(ANOVA)
2 signals
PoP pair forecast
Individual
Forecasts
(ARIMA)
Weekly time series
8
www.intel.com/research
• Intel Research •
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
(2^6*1.5 hours=96 hours) using B3 spline filter.
p
x (t )  c p (t )   d j (t )
j 1
d j (t )  c j 1 (t )  c j (t )
9
www.intel.com/research
• Intel Research •
Wavelet Decomposition
Approximations
Details
10
www.intel.com/research
• Intel Research •
Methodology Roadmap
SNMP
Topology
Traffic
Aggregates
1 signal
Wavelet
Multiresolution
Analysis
7 signals
Model
Reduction
(ANOVA)
2 signals
PoP pair forecast
Individual
Forecasts
(ARIMA)
Weekly time series
11
www.intel.com/research
• Intel Research •
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)
12
www.intel.com/research
• Intel Research •
Analysis of Variance
x(t )  c6 (t )    d3 (t ),
 3
 Accounts for 80-94% of total variance
 Time series can be easily further compacted into
weekly time series
13
www.intel.com/research
• Intel Research •
Methodology Roadmap
SNMP
Topology
Traffic
Aggregates
1 signal
Wavelet
Multiresolution
Analysis
7 signals
Model
Reduction
(ANOVA)
2 signals
PoP pair forecast
Individual
Forecasts
(ARIMA)
Weekly time series
14
www.intel.com/research
• Intel Research •
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


x(l )  l ( j)    dˆt3 (l ),
 3
15
www.intel.com/research
• Intel Research •
Evaluation of Forecasts
16
www.intel.com/research
• Intel Research •
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.
17
www.intel.com/research
• Intel Research •
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?
18
www.intel.com/research
• Intel Research •
Questions?
Thank you!
19
www.intel.com/research