Title: Traffic Matrix Estimation in Non-Stationary Environments
1Traffic Matrix Estimation in Non-Stationary
Environments
- Presented by
- R. L. Cruz
- Department of Electrical Computer Engineering
- University of California, San Diego
- Joint work with
- Antonio Nucci
- Nina Taft
- Christophe Diot
- NISS Affiliates Technology Day on Internet
Tomography - March 28, 2003
2The Traffic Matrix Estimation Problem
- Formulated in Y. Vardi, Network Tomography
Estimating Source-Destination Traffic From Link
Data, JASA, March 1995, Vol. 91, No. 433, Theory
Methods
3The Traffic Matrix Estimation Problem
Xj
Yi
ingress
egress
Xj
PoP (Point of Presence)
Y A X
Traffic Matrix
Link Measurement Vector
Routing Matrix
4The Traffic Matrix Estimation Problem
- Importance of Problem capacity planning, routing
protocol configuration, load balancing policies,
failover strategies, etc. - Difficulties in Practice
- missing data
- synchronization of measurements (SNMP)
- Non-Stationarity (our focus here)
- long convergence time needed to obtain estimates
5What is Non-Stationary?
- Traffic Itself is Non-Stationary
6What is Non-Stationary?
- Also, Routing is Non-Stationary
- e.g. Due to Link Failures
- Essence of Our Approach
- Purposely reconfigure routing in order to help
estimate traffic matrix - More information leads to more accurate estimates
- Effectively increases rank of A
- We have developed algorithms to reconfigure the
routing for this purpose (beyond the scope of
this talk)
7Outline of Remainder of Talk
- Describe the Stationary Method
- Stationary traffic, non-stationary routing
- Stationary traffic assumption is reasonable if we
always measure traffic at the same time of day
(e.g. peak period of a work day) - Briefly Describe the Non-Stationary Method
- Both non-stationary traffic and non-stationary
routing - More complex but allows estimates to be obtained
much faster
8Network and Measurement Model
- Network with L links, N nodes, PN(N-1) OD pair
flows - K measurement intervals, 1 k K
- Y(k) is the link count vector at time k (L x 1)
- A(k) is the routing matrix at time k (L x P)
- X(k) is the O-D pair traffic vector at time k (P
x 1) - X(k) (x1(k) , x2(k) , xP(k))T
Y(k) A(k) X(k)
Y(k) and A(k) can be truncated to reflect missing
and redundant data
9Traffic Model Stationary Case
- X(k) is the O-D pair traffic vector at time k
(P x 1) - X(k) (x1(k) , x2(k) , xP(k))T
X(k) X W(k)
- W(k) Traffic Fluctuation Vector
- Zero mean, covariance matrix B
- B diag(X)
10Matrix Notation
Linear system of equations
where
LK
LKP
LKKP
KP
P
Choose Routing Configurations such that
Rank(A) P
11Traffic matrix Estimation-Stationary Case
Y AX CW
- Initial Estimate Use Psuedo-Inverse of A-
does not require statistics of W (covariance B) - Gauss-Markov Theorem Assume B is known
- - Unbiased, minimum variance estimate-
Coincides with Maximum Likelihood Estimate - if W is Gaussian
12Traffic matrix Estimation-Stationary Case
Y AX CW
- Minimum Estimation Error (assumes B is
known)
13Traffic matrix Estimation-Stationary Case
- Recall we assume B cov(W) satisfies B
diag(X) - Set
14Traffic matrix Estimation-Stationary Case
- Our estimate is a solution to the equation
- Open questions for fixed point equation
- Existence of Solution?- Uniqueness?
- Is solution an un-biased estimate?
15Numerical Example-Stationary case
- N10 nodes, L24 links and P90 connections.
- Three set of OD pairs with mean x equal to
- 500 Mbps, 2 Gbps and 4 Gbps.
- Gaussian Traffic Fluctuations
16Stationary case b1 Samples/Snapshot1
17Stationary case b1 Samples/Snapshot1
18Stationary case b1 Samples/Snapshot15
19Stationary case b1 Samples/Snapshot15
20Stationary case b1.4 Samples/Snapshot1
21Stationary case b1.4 Samples/Snapshot1
22Stationary case b1.4 Samples/Snapshot15
23Stationary case b1.4 Samples/Snapshot15
24Stationary and Non-Stationary traffic
- 20 snapshots / 4 samples per snapshot / 5 min per
sample - Stationary Approach 20 min per day (same time)
/ 20 days - Non-Stationary Approach aggregate all the
samples in - one window time large 400 min (7 hours)
25Traffic Model Non-Stationary Case
- Each OD pair is cyclo-stationary
- Each OD pair is modeled as
- Fourier series expansion
26Mean estimation Results-Non Stationary case
- Three set of OD pairs
-
- where are linear independent Gaussian
variables with
27Non Stationary case b1 Link Count
28Non Stationary case b1 Mean estimation