Proposals on the Source-Destination Trac Matrix

Estimation for IP-Based VPNs

Shigeo Shioda and Kazuya Ohtani
Department of Urban Environment and Systems, Faculty of Engineering, Chiba University
1-33 Yayoi, Inage, Chiba 2638522, Japan
Email: [email protected]
AbstractWe study the problem of estimating a source-
destination (SD) trac matrix of an IP-based VPN from mea-
surements of aggregated trac along all access links. We derive
two dierent estimators by formulating the SD-trac-matrix
estimation as a deterministic or stochastic optimization problem.
One estimator does not rely on any stochastic information
about the SD trac and the other estimator assumes the linear
relationship between the average and variance of the SD trac.
The proposed estimators have expressions much simpler than
those of standard SD-trac-matrix estimators; one estimator has
an explicit representation that does not involve matrix inversion.
The other is given by a simple matrix formula and does not need
any prior SD-trac matrices. Through simulation experiments,
we show that these proposals are superior or comparable to the
standard SD-trac-matrix estimator in term of the estimation
accuracy.
I. Ixraootcrrox
The progress in IP technologies has accelerated the use of
IP-based virtual private networks (IP-based VPNs) in recent
years. Network service providers have oered various IP-
VPN services, which have enabled customers to construct IP-
based VPNs on a public network infrastructure with a security
guarantee comparable to networks constructed with private
lines. IP-VPN service is less expensive than the conventional
private-line service for constructing a private network covering
many network sites. Thus, it appeals to large companies that
have a lot of branch oces in dierent locations.
Since the number of endpoints per VPN is growing, trac
patterns between endpoints are becoming increasingly dicult
to forecast. Thus, IP-VPN-service customers will generally not
contract for bandwidths between all source-destination (SD)
pairs of endpoints. Rather, they will contract only for the
bandwidths of access links between VPN sites and egresses
(edge routers) of the providers network. With such a form of
contract, however, it is dicult to guarantee the Quality-of-
Service (QoS) between all SD-endpoint pairs of a VPN. (The
QoS-support model, where only the bandwidths between VPN
sites and egresses of the providers network are guaranteed, is
called the hose model [1].)
Recently, Dueld et al. [2] proposed a new VPN ser-
vice that oers QoS-guaranteed VPNs to customers who are
unwilling (or unable) to specify the detailed trac patterns
between endpoints. In this service, bandwidth and thereby
the QoS is guaranteed between all SD-endpoint pairs within
the providers network, and the bandwidth is dynamically
allocated based on online-trac measurement. To implement
such a service, the network service provider needs to know
the tracs (the amounts of data in bit transferred in unit
time) between all source-destination pairs of endpoints of the
VPN. Knowing the SD-trac matrix is also very valuable
to a wide variety of trac engineering tasks including load
balancing, routing optimization, pricing, capacity planning,
and trouble shooting [3]. For a typical network, however, direct
measurement of the SD-trac matrix is not usually feasible,
but less informative measurements of aggregated trac along
the link of the network are much easier to perform [4].
In this context, we consider the problem of estimating the
SD-trac matrix from measurements of aggregated trac
along all access links between VPN sites and egresses of
the providers network. We derive two dierent estimators by
formulating the estimation of the SD-trac matrix as both
deterministic and stochastic optimization problems.
The remainder of this paper is organized as follows. Sec-
tion II summarizes related work on the SD-trac estimation
and states the target of this paper. Section III introduces several
denitions and the notation used in this paper. Section IV dis-
cusses the estimation of the SD-trac matrix in a deterministic
setting and derives the estimator that has explicit representa-
tion not involving matrix inversion. Section V discusses the
problem in a stochastic setting and derives the other estimator
under the assumption that there exists a linear relationship
between average and variance of the SD trac. Section VI
provides numerical results to compare the estimator by Cao
[5], which is recognized a standard technique for SD trac
estimation, with respect to the accuracy of estimation. We also
show the sensitivity of the proposed estimator to some prior
information or assumption.
II. Rrixrro woak xxo ota coxraratrrox
The estimation of the amounts of trac between SD-
endpoint pairs was rst discussed in the eld of telephone
network planning. The goal of the estimation is to identify
the amounts of trac (or trac intensities) between SD pairs
of local exchanges based on observed data about aggregated
trac originating from and terminating at all local exchanges
(that is, when a telephone network has N local exchanges,
there are N
2
trac data to be estimated and 2N trac data
available for estimation). Since N
2
is larger than 2N for
N 3, there exist many possible SD-trac patterns that
matter experts for publication in the IEEE GLOBECOM 2005 proceedings. This full text paper was peer reviewed at the direction of IEEE Communications Society subject
IEEE Globecom 2005 348 0-7803-9415-1/05/$20.00 2005 IEEE
reproduce the observed data about aggregated trac. Thus, the
estimation problem is equivalent to nding the most likely
pattern that reproduces the observed trac data. Saito and
Abe [6] formulated the SD-trac estimation as a deterministic
optimization problem, which aims to nd the closest SD-
trac pattern to a prior pattern among all possible patterns
that reproduce the observed data about aggregated trac, and
derived a simple solution for this problem. The problem setting
of SD-trac estimation in telephone networks is very similar
to the problem considered in this paper.
Since it aects the design, routing, conguration, and pric-
ing of computer networks, especially the Internet, SD-trac
estimation has attracted a great deal of attention recently. The
proposals by Vardi [4] and by Cao et al. [5] used statistical
inference techniques: under the assumption that the trac is
generated by some probability distributions, they estimate the
parameters (for example, the mean amount of trac) of the
probability distributions [5], [4]. Medina et al. [3] made a
detailed comparative evaluation of these existing techniques
and proposed to incorporate gravity (or choice) model with
these existing proposals to increase the accuracy. Zhang et
al. [7] also discusses the use of gravity model in SD-trac-
matrix estimation. Roughan et al. [8] studied the impact of
the accuracy of the estimated trac matrix on the routing
optimization.
Note that the common formulation for the SD-trac estima-
tion discussed by these recent publications diers somewhat
from the problem discussed in the telephone network and our
problem setting. In the former problem, the amounts of trac
are counted on all links of the network and all of these are
used for the SD-trac estimation. In the latter problem (SD
trac estimation in telephone network or our problem), the
amounts of trac are counted only on the links originating
from and terminating at all local exchanges (VPN sites). Thus,
in the latter problem, the estimation should be conducted under
much less information than the former.
The aim of this paper is to derive practical and easy-to-
use estimators of the SD-trac matrix in the problem setting
where only aggregated trac originating from and terminating
at local-VPN sites can be observed. The main contribution of
this paper is to show that, in such a problem setting, we can
have estimators much simpler than those obtained by existing
techniques [4], [5], [3], [8] proposed for computer networks.
For example, one estimator proposed in this paper has explicit
representation that does not involve matrix inversion. The other
estimator is given by a simple formula, which works based
on the information that is observable from aggregated trac
along access links. These features make signicant contract
to the existing estimators: they usually need very complicated
calculation or some prior information about SD trac matrix.
III. Drrrxrrroxs
Letting N be the number of endpoints of the VPN, we dene
(Fig. 1):
- x
(i1)N+j
: the trac from the ith endpoint to the jth
endpoint,
- y
i
(1 i N): the trac to the ith endpoint out of the
VPN,
- y
i+N
(1 i N): the trac from the ith endpoint into the
VPN,
where the trac means the amount of data (in bit or byte)
transferred in unit time (say a few minutes). In what follows,
Providers
Network
ith endpoint
jth VPN sites
i
y
ith VPN sites
N i
y
+
j N i
x
+ ) 1 (
jth endpoint
Fig. 1. VPN model and notation
vector x = (x
1
, , x
N
2 )
T
denotes SD tracs and vector
y = (y
1
, , y
2N1
)
T
denotes access-link tracs. These vectors
satisfy the relationship
y = Ax, (1)
where A is an (2N 1) N
2
matrix given by
A =
,

I
N
I
N
I
N
I
N
e
T
N
0
T
N
0
T
N
0
T
N
0
T
N
e
T
N
0
T
N
0
T
N
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
T
N
0
T
N
e
T
N
0
T
N

,
where I
N
is an N N identity matrix, e
N
is an N-dimensional
column vector with all elements equal to 1, and 0
N
is an N-
dimensional column vector with all elements equal to 0. Our
goal is to estimate the unobserved SD trac vector x from the
observed access-link trac vector y. Although in some cases
we use the customary terminology trac matrix, we arrange
the tracs of all SD pairs in a single vector x for convenience
as in [5].
Remark 1: It seems more natural to dene y and A as
follows:
y = (y
1
, , y
2N
)
T
,
A =
,

I
N
I
N
I
N
I
N
e
T
N
0
T
N
0
T
N
0
T
N
0
T
N
e
T
N
0
T
N
0
T
N
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
0
T
N
0
T
N
e
T
N
0
T
N
0
T
N
0
T
N
0
T
N
e
T
N

.
In the above denitions, however, y = Ax consists of 2N linear
equations that are not independent of each other because one
of them is redundant. So, we do not use these denitions.
matter experts for publication in the IEEE GLOBECOM 2005 proceedings. This full text paper was peer reviewed at the direction of IEEE Communications Society subject
IEEE Globecom 2005 349 0-7803-9415-1/05/$20.00 2005 IEEE
IV. Drrramrxrsrrc raoairm
A. Formulation
First, we formulate the estimation of the SD-trac matrix as
a deterministic problem. Let x be the estimate of x for a given
y. Since (1) consists of 2N 1 linear equations and x has N
2
elements, the information provided by the access-link trac
y is insucient to determine a unique SD-trac vector from
(1). Thus, we impose an additional constraint: the estimate
x should be the closest vector to x
0
among all solutions of
(1), where x
0
is a vector representing some prior SD trac
vector. That is, x is given by the solution of the following
optimization problem:
Minimize | x x
0
|, subject to y = Ax. (2)
B. Estimator via generalized inverse matrix
It is shown in [9] that the solution of (2) is given by
x(y; x
0
) = x
0
+ A
+
(y Ax
0
), (3)
where A
+
is Moore-Penroses generalized inverse matrix [9].
If the number of rows of A is equal to rank(A), then A
+
is
equal to A
T
(AA
T
)
1
, where A
T
is the transform of A. Thanks
to the special structure of A, which owes to the problem setting
(only the amounts of aggregated trac originating from and
terminating at local VPN-sites are observable) in this paper,
the inverse of AA
T
and thus A
+
has the explicit representation
such as
A
+
=
1
N
,

I
N
E
N,N
/N [e
N
, 0
N
, , 0
N
]
I
N
E
N,N
/N [0
N
, e
N
, , 0
N
]
.
.
.
.
.
.
I
N
E
N,N
/N [0
N
, 0
N
, , e
N
]
I
N
+ (N 1)E
N,N
/N E
N,N1

,
where E
m,n
is an mn matrix with all elements equal to one.
The estimate of the trac from the mth endpoint to the nth
endpoint, x
(m1)N+n
, is thus given by
x
(m1)N+n
= x
0
(m1)N+n
+
1
N

n

1
N
N

k=1

k
+
N+m

for m < N, (4)

and
x
(N1)N+n
= x
0
(N1)N+n
+
1
N

n
+
N 1
N
N

k=1

k

N1

k=1

N+k

, (5)
where
n
is the nth element of vector given by
= y Ax
0
.
The explicit representation of the estimator such as (4) and
(5) is not usually available in more general problem setting of
SD-trac matrix estimation.
The performance of the estimator should depend on the
prior information x
0
. Here, we show an interesting nding
concerning the choice of prior. To explain this, we assume the
following prior vector:
x
0
= e
N
2 , (6)
where denotes some constant value. We call (6) the uniform
prior in this paper. We have found that using the uniform prior
is equivalent to choosing the prior at origin (that is, x
0
= 0)
and that (4) and (5) with uniform prior are comparable with
the estimator by Cao [5] in terms of estimation accuracy.
V. Srocnxsrrc raoairm
Next, we formulate the estimation of SD-trac matrix as a
stochastic problem. In this section, let X = (X
1
, , X
N
2 ) be
a random vector that represents SD-trac vector, and let Y =
(Y
1
, , Y
2N1
) be a random vector that represents the access-
link trac (using capital letters to distinguish the stochastic
case). Furthermore, we let be the expectation of X and
xx
=
E[(X )(X )
T
] be the variance matrix of X. We assume
that the following linear relationship between the average and
variance of the SD trac exists:
Var[X] = . (7)
Similar relationships between the average and variance of the
SD trac have been frequently used in recent publications on
SD-trac-matrix estimation [5], [3].
Our target in the stochastic problem setting is to nd the
most probable x (an outcome of X), which satises (1), from
y (an outcome of Y), , and
xx
. In this section, we show that
the derived estimator does not require the knowledge about .
Then we show how the variance matrix
xx
can be estimated
from repeated measurements of Y.
A. Formulation and solution
We assume that X follows a normal distribution whose
probability density function (pdf) is given by
f (x|,
xx
)
def
= exp{
1
2
(x )
T

1
xx
(x )}.
Our goal is to nd the estimator, among all solutions of (1),
that maximizes the above pdf. That is, the estimator x is given
by the solution to the following optimization problem:
Maximize f (x|,
xx
), subject to y = Ax. (8)
Mathematical manipulation frequently used in a stochastic
inference problem leads to the following solution:
x = +
xx
A
T
(A
xx
A
T
)
1
(y A), (9)
where, under assumption (7), we can show that the following
relationship holds:

xx
A
T
(A
xx
A
T
)
1
A = . (10)
Thus, we nally have
x =
xx
A
T
(A
xx
A
T
)
1
y. (11)
Surprisingly, the above estimator does not require the knowl-
edge about but just only
xx
. Note that the variance matrix

xx
is observable from Y as shown in next subsection. This
means that estimator (11) can work without any prior SD-
trac matrix. This feature makes signicant contrast to the
most existing proposals on SD-trac estimation. We also note
matter experts for publication in the IEEE GLOBECOM 2005 proceedings. This full text paper was peer reviewed at the direction of IEEE Communications Society subject
IEEE Globecom 2005 350 0-7803-9415-1/05/$20.00 2005 IEEE
that the simple representation of estimator (11) owes to the
problem setting (in other words, the special structure of A)
assumed in this paper, and it is not available for more general
SD-trac estimation.
The performance of the estimator should depend on whether
assumption (7) holds or not. This dependency is numerically
studied in Section VI.
Remark 2: The linear-least-squares-error estimator is also
given by formula (9), where the assumption that X follows a
normal distribution is not required.
Remark 3: Estimator (11) is a natural extension of estima-
tor (3) in the sense that estimator (11) gives the closest
pattern of SD-trac intensity to the zero vector 0 in the space
where the distance between vectors x
1
and x
2
is dened by

(x
1
x
2
)
T

1
xx
(x
1
x
2
).
B. Estimation of the variance matrix of X
We can estimate the variance matrix of X,
xx
, from
repeated measurements of Y under the assumption that the
elements of X are statistically independent of each other. To
show this, we start by noting that the nondiagonal elements
of
xx
are all equal to zero because of the independence
assumption. Next, note that the following relationship between
the elements of X and Y holds:
Y
i
=
N

k=1
X
(k1)N+I
, Y
N+j
=
N

l=1
X
( j1)N+l
.
Thus, the independence assumption yields
Cov[Y
i
, Y
N+j
] = Cov[
N

k=1
X
(k1)N+I
,
N

l=1
X
( j1)N+l
]
= Var[X
( j1)N+i
].
Let y
(1)
, , y
(m)
be a consecutive set of measurements of Y.
The above relationship enables us to use the following formula
to estimate the variance of SD trac:
Var[X
( j1)N+i
]
1
m
m

k=1

y
(k)
i

1
m
m

l=1
y
(l)
i

y
(k)
N+j

1
m
m

l=1
y
(l)
N+j

,
(12)
which gives us estimates of the diagonal elements of
xx
.
VI. Srmtixrrox rxrrarmrxrs
We evaluated our proposals through simulation experiments
with respect to the estimation accuracy and sensitivity to some
prior information and assumption. We considered two VPN
models: ve-endpoint and seven-endpoint models. To simulate
the network, we used synthetic average SD-trac matrices.
The average SD trac of the ve-endpoints network ranges
from 0 Mbps to 8 Mbps (Table I), and that of the nine-
endpoints network ranges from 5 Mbps to 50 Mbps (Table
II). Based on each of these average SD-trac matrices,
we randomly generated daily SD-trac matrix for 1000
days, x
(1)
, , x
(1000)
, according to the normal distribution,
and produced daily access-link trac vectors for 1000 days,
y
(1)
, , y
(1000)
. We then used our proposals to obtain the
TABLE I
Avraxor SD-raxrrrc mxrarx roa rrvr-rxororxr xrrwoak (rx Mars)
Destination
5 4 3 2 1
0.07 0.04 0.95 3.57 7.91 5
1.33 1.18 2.92 0.73 0.12 4
2.43 4.11 5.68 6.87 1.39 3
5.97 6.58 7.17 2.80 3.84 2
4.68 6.47 1.55 4.51 0.01 1
Origin
TABLE II
Avraxor SD-raxrrrc mxrarx roa srvrx-rxororxr xrrwoak (rx Mars)
Origin
7
6
5
4
3
2
1
Destination
7 6 5 4 3 2 1
43.8 15.6 25.8 11.4 29.3 46.7 48.0
37.7 44.4 18.6 28.4 41.1 40.3 32.4
7.57 20.9 25.3 34.8 12.5 32.3 32.1
30.7 28.9 22.0 5.40 5.21 10.4 25.1
49.5 12.5 11.6 21.4 9.11 5.67 18.7
28.1 37.0 43.7 12.8 38.6 42.0 45.3
20.8 26.6 31.3 41.4 13.7 30.4 5.06
estimate, x
(1)
, , x
(1000)
, and evaluated the relative estimation
error
e
(n)
def
=
|x
(n)
x
(n)
|
|x
(n)
|
, n = 1, , 1000,
for 1000 days. To generate the daily SD-trac matrix, we
assumed that the mean and variance are related according to
the relationship
Var[X] =
c
, (13)
which has been used by Cao et al. [5] and Medina et al. [3].
A. Accuracy of estimation
First, we compared the proposed estimators with the esti-
mator by Cao [5] in terms of accuracy of estimation. Note that
the estimator [5] is considered as the standard technique for
SD-trac estimation [3]. The results are summarized in Table
III, where Estimator A denotes the estimation based on (4) and
(5), and Estimator B denotes the estimation based on (11). All
of the results were obtained by generating the daily SD-trac
matrices by setting = 0.5 and c = 1 in (13). We used the
uniform prior in Estimator A ((4) and (5)). The variance matrix

xx
in Estimator B was obtained by (12) based on the daily
access-link trac vectors for 1000 days, y
(1)
, , y
(1000)
.
Table III indicates that, Estimator with the uniform prior
is comparable with the proposal by Cap [5] in term of the
estimation accuracy. The table also indicates that Estimator B
signicantly outperforms Estimator A and the estimator by
[5]. Similar results were obtained with other choice of trac
matrices. These results are not surprising because the proposed
matter experts for publication in the IEEE GLOBECOM 2005 proceedings. This full text paper was peer reviewed at the direction of IEEE Communications Society subject
IEEE Globecom 2005 351 0-7803-9415-1/05/$20.00 2005 IEEE
TABLE III
Acctaxc. or rsrrmxrrox: cxrxarsox wrrn Cxos raorosxi
Estimator A
67.9
90% error
in prior
23.1
30% error
in prior
9.1 Estimator B
49.4 Estimator by Cao
49.5
Estimator A with
the uniform prior
53.4 38.2 8.15
70% error
in prior
50% error
in prior
10% error
in prior
Relative estimation error [%]
TABLE IV
Srxsrrrvrr. ro rnr raroa. Frvr-rxororxr xrrwoak
Estimator A
67.9
90% error
in prior
23.1
30% error
in prior
9.1 Estimator B
49.4 Estimator by Cao
49.5
Estimator A with
the uniform prior
53.4 38.2 8.15
70% error
in prior
50% error
in prior
10% error
in prior
Relative estimation error [%]
TABLE V
Srxsrrrvrr. ro rnr raroa. Srvrx-rxororxr xrrwoak
Estimator A
79.0
90% error
in prior
25.3
30% error
in prior
21.2 Estimator B
40.9 Estimator by Cao
36.1
Estimator A with
the uniform prior
60.0 42.1 8.46
70% error
in prior
50% error
in prior
10% error
in prior
Relative estimation error [%]
estimators are tailored to the problem setting considered in this
paper, while the estimator by Cao is not.
B. Sensitivity to the prior SD-trac matrix
The performance of Estimator A depends on the prior
SD-trac matrix x
0
. Thus, we evaluated the sensitivity of
Estimator A to the prior SD-trac matrix. For this purpose, we
generated the prior SD-trac vector x
0
by adding white noise
to the mean SD-trac-intensity vector; that is, x
0
= + ,
where is a parameter to adjust the error size and is the
white noise.
Tables IV and V respectively show the average relative
estimation error for ve- and seven-endpoint networks. In the
tables, the error in prior SD trac column denotes
|

x
0
|
||
multiplied by 100 to get a percentage. All of the results were
obtained by generating the daily SD-trac matrices by setting
= 0.5 and c = 1 in (13).
These tables indicate that Estimator A with good prior
outperforms Estimator B and Caos estimator. The variance
matrix in (11), which was measured by the method explained
in Section V-B, should contain some measurement error,
which gives the reason why Estimator A with good prior
outperform Estimator B. In addition to this, we see that using
the uniform prior corresponds to using the prior with an error
of around 50%.
C. Sensitivity to the error in variance matrix
As explained in VI-B, the variance matrix in (11) inevitably
contains a measurement error. Thus, we evaluated sensitivity
of Estimator B to the error in variance matrix by the following
procedure: we produced an incorrect variance matrix

xx
by
adding white noise to the diagonal elements of the original
variance matrix
xx
. We then used the incorrect variance
matrix

xx
in estimation. Experiments by simulation were also
conducted with = 0.5 and c = 1 in (13).
TABLE VI
Srxsrrrvrr. ro
xx
. Frvr-rxororxr xrrwoak.
49.5 Estimator A
Estimator B
101.1
90% error
in variance
22.1
30% error
in variance
9.1
Estimator B with
measured variance
49.4 Estimator by Cao
83.2 42.3 7.17
70% error
in variance
50% error
in variance
10% error
in variance
Relative estimation error [%]
TABLE VII
Srxsrrrvrr. ro
xx
. Srvrx-rxororxr xrrwoak.
36.1 Estimator A
Estimator B
99.4
90% error
in variance
26.9
30% error
in variance
21.2
Estimator B with
measured variance
36.1 Estimator by Cao
82.4 53.2 8.15
70% error
in variance
50% error
in variance
10% error
in variance
Relative estimation error [%]
The relative estimation errors are given in Tables VI (ve-
endpoint network) and VII (seven-endpoint network). The rst
point to note is that Estimator B is highly sensitive to the vari-
ance information. The second point to note is that Estimator B
does not perform better than Estimator A or the estimator by
Cao when there is a lot of noise in the variance information.
For example, for the ve-endpoint network, Estimator B does
not produce better results than Estimator A when the variance
contains an error of more than 70%. For the seven-endpoint
network, variance information with an error of 50% or more
makes Estimator B perform worse than Estimator A.
For reference, we show the relative estimation error of
Estimator B when the variance matrix was measured based
on the 1000-days observation of access link loads. The table
indicates that the 1000-days observation of access link loads
yields the estimate of the variance matrix with an error of
matter experts for publication in the IEEE GLOBECOM 2005 proceedings. This full text paper was peer reviewed at the direction of IEEE Communications Society subject
IEEE Globecom 2005 352 0-7803-9415-1/05/$20.00 2005 IEEE
less than 30%. Through other numerical experiments, we
have conrmed that, for the seven-endpoint network, 100-days
observation of access link loads is required to produce the
estimate of the variance matrix with an error of less than 50%.
D. Sensitivity to the relationship between the mean and vari-
ance
Finally, we examined the sensitivity to the linear relation-
ship between the mean and variance. Assuming the relation-
ship (13) by setting at 0.5, we varied the exponent c in the
range from 0.5 to 2 to examine the sensitivity to c. A recent
study [10] revealed that the actual trac certainly satises
the mean-and-variance relationship 13 and the exponent c is
around 1.5. The results are summarized in Fig. 2 (ve-endpoint
network) and Fig. 3 (seven-endpoint network). The prior SD-
trac matrix in Estimator A ((4) and (5)) used the uniform
prior.
R
e
l
a
t
i
v
e

e
s
t
i
m
a
t
i
o
n

e
r
r
o
r
c
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.5 1 1.5 2
Estimator B
Estimator A
Estimator by Cao
Fig. 2. Sensitivity to the mean and variance relationship. Five endpoint
network.
R
e
l
a
t
i
v
e

e
s
t
i
m
a
t
i
o
n

e
r
r
o
r
c
0
0.1
0.2
0.3
0.4
0.5
0.6
0.5 1 1.5 2
Estimator B
Estimator A
Estimator by Cao
Fig. 3. Sensitivity to the mean and variance relationship. Seven endpoint
network.
These tables indicate that the proposed estimators (Esti-
mators A and B) have some dependence on c while the
estimator by Cao does not. The tables, however, also indicate
that Estimator A always yields estimates comparable to those
by Caos estimator regardless of the exponent. In addition to
this, in most cases, Estimator B performs better than both
Estimator A and Caos estimator regardless of the exponent.
That is, although Estimator B relies on the linear relationship
between the mean and variance (7), it still performs well even
if relationship (7) does not hold.
VII. Coxcitorxo armxaks
We discussed the problem of estimating the trac intensities
between all SD-endpoint pairs of a VPN from measurements
of trac along access links. We proposed two estimators: the
rst estimator ((4) and (5); Estimator A) does not require any
statistical information about trac and has explicit representa-
tion that does not have matrix inversion. The second estimator
((11); Estimator B), which assumes the linear relationship
between average and variance of SD trac, is also given by
a simple matrix formula and does not require any prior SD-
trac matrices.
In this study, we found that Estimator A with the uniform
prior is comparable to the estimator by Cao in terms of the
estimation accuracy. This is a very remarkable nding because
Estimator A is much easier to compute than the estimator by
Cao. We also found that Estimator B is very accurate if the
variance matrix
xx
can be estimated with a small error. We
believe that Estimator B is promising for small-size networks
because the variance matrix
xx
of small network would be
precisely estimated from a modest number of measurements
of access link loads. For more general networks, however,
Estimator A seems to be useful.
Some recent studies [7] indicated that most of trac matri-
ces approximately agree with the gravity model and the use of
the gravity model could largely improve the accuracy of the
estimation. We have found that Estimator A with the gravity-
model prior yields very accurate results when the trac matrix
(approximately) satises the gravity model. We also found
that the gravity model can be used to estimate the variance
matrix in Estimator B, which yields explicit representation
of Estimator B like Estimator A. These ndings have been
reported in [11].
Rrrrarxcrs
[1] V. Davie and Y. Rekhter, MPLS: Technology and Applications, Aca-
demic Press, 2000.
[2] N.G. Dueld et al., A exible model for resource management in
virtual private networks, in ACM SIGCOMM 99, 2002, pp. 95108.
[3] A. Medina, N. Taft, K. Salamatian, S. Bhattacharyya, and C. Diot,
Trac matrix estimation: existing techniques and new directions, in
ACM SIGCOMM 02, 2001, pp. 161174.
[4] Y. Vardi, Network tomography: estimating source-destination trac
intensities from link data, Journal of the American Statistical Associ-
ation, vol. 91, no. 433, pp. 365377, 1996.
[5] J. Cao, D. Davis, S.V. Wiel, and B. Yu, Time-varying network
tomography: router link data, Journal of the American Statistical
Association, vol. 95, no. 452, pp. 10631075, 2000.
[6] H. Saito and T. Abe, Methods of estimating trac matrices, IEICE
Trans. Commun., vol. J69-B, no. 10, pp. 10311037, 1986.
[7] Y. Zhang, M. Roughan, N.G. Dield, and A. Greenberg, Fast accurate
computation of large-scale IP trac matrices form link loads, in ACM
SIGMETRICS 03, 2003, pp. 206217.
[8] M. Roughan, M. Thorup, and Y. Zhang, Performance of estimated
trac matrices in trac engineering, in ACM SIGMETRICS 03, 2003,
pp. 326327.
[9] C.R. Rao and S.K. Mitra, Generalized Inverse of Matrices and its
Applications, Wiley, 1971.
[10] A. Gunnar, M. Johansson, and T. Telkamp, Trac matrix estimation
on a large IP backbone, in ACM Internet Measurement Conference,
2004, pp. 149160.
[11] S. Shioda and K. Ohtani, Estimating the source-destination trac
matrix of a VPN from access-link loads, in International Conference
on Telecommunication Systems - Modeling and Analysis, 2005.
matter experts for publication in the IEEE GLOBECOM 2005 proceedings. This full text paper was peer reviewed at the direction of IEEE Communications Society subject
IEEE Globecom 2005 353 0-7803-9415-1/05/$20.00 2005 IEEE