Professional Documents
Culture Documents
Queueing Analysis of High-Speed Multiplexers Including Long-Range Dependent Arrival Processes
Queueing Analysis of High-Speed Multiplexers Including Long-Range Dependent Arrival Processes
Queueing Analysis of High-Speed Multiplexers Including Long-Range Dependent Arrival Processes
t
t
Fig. 1. A uid queueing system with an innite buffer and a server.
= P
__
sup
ts0
(N
0
N
s
) > x
__
t
P
__
sup
s0
(N
0
N
s
) > x
__
. (1)
Hence, P({Q > x}) := lim
t
P({Q
t
> x}) =
P
__
sup
t0
(N
0
N
t
) > x
__
. So if we dene X
t
:= N
0
N
t
, then {X
t
: t 0} is a stochastic process with station-
ary increments, and P({Q > x}) = P
__
sup
t0
X
t
> x
__
.
For notational simplicity, henceforth, we dene w
:=
sup
However,
x,
even though it is a very different function of x. Therefore,
approximations for P({X > x}), based on (3), should be
used with some caution, since (3) provides relatively weak
theoretical support to the accuracy of these approximations.
In order to address this difculty, our objective in this paper
is to focus on Gaussian processes, including many types of
long-range dependent processes, and develop a considerably
stronger asymptotic relation.
Recently, Gaussian processes have received a lot of attention
for the modeling and analysis of queueing behavior in high-
speed networks [1], [5], [7], [8], [17]. There are many reasons
for this. Due to the huge link capacity of high-speed networks,
hundreds or even thousands of network applications are likely
to be served by a multiplexer. For example, an OC 3 line
(155.52 Mbps) can accommodate over 7700 typical voice calls
at a link utilization of 0.8. An OC 12 line (622.08 Mbps)
can accommodate over 300 MPEG-1 (1.5 Mbps) video calls
at the same link utilization. Many companies sell commercial
switches that support OC 12 lines, and ATM networks with
OC 24 (1.2 Gbps) lines are already operational (at Cam-
bridge University, for example). Also, switches supporting
link capacities of several gigabits-per-second (and higher) are
on the horizon, all of which suggests that by appealing to the
Central Limit Theorem, we can accurately characterize the in-
put process as a Gaussian process.
In earlier studies with trafc modeling we have empirically found that, typ-
ically, a couple of hundred multiplexed sources are sufcient for the trafc to
be modeled as a Gaussian process [5], [7].
is Gaussian, P({X
t
> x}) can be expressed in terms of
v
t
, , and the standard Gaussian tail function (w) :=
_
w
exp
_
z
2
/2
dz/
2, as
P({X
t
> x}) =
_
x +t
v
t
_
.
Assuming that v
t
/ t
2
0 as t , it is not difcult to see
that v
t
_
(x +t)
2
should attain its maximumvalue at some -
nite t = t
x
, and therefore, the probability P({X
t
> x}) is also
maximized at t = t
x
. The qualitative statement rare event
take place only in the most probable way (e.g. see [10]) sug-
gests that P({X
tx
> x}) = sup
t
P({X
t
> x}) should be
a good lower bound approximation for P({X > x}), and
in fact similar ideas have already been used in different ways
to analyze and approximate the tail probability [7], [10], [11].
In [7], [8], we have provided a rigorous asymptotic result that
theoretically supports the above qualitative statement (this re-
sult is both generalized and strengthened by Theorem 3 in this
paper). Also, in those papers, for a fairly large class of Gaus-
sian processes where the tail probability is asymptotically ex-
ponential, we have found that (i) m
x
:= (x +t
x
)
2
/v
tx
(the
reciprocal of the maximum value of v
t
_
(x +t)
2
) contains
important information about the shape of the tail probability
curve, (ii) exp[m
x
/ 2] asymptotically bounds the tail proba-
bility from above, and (iii) exp[m
x
/ 2] provides a very accu-
rate estimate of the tail probability over a wide range of queue
lengths x, including small values of x.
In this paper, we consider a more general class of Gaussian
processes, including a large class of long-range dependent pro-
cesses, to show that
log P({X > x}) +
m
x
2
O(log x), (4)
where O(f(x)) denotes the set of functions g(x) such that
limsup
x
| g(x)/ f(x)| < . Observe that (4) character-
izes log P({X > x}) in much more detail than (3). Further,
(4) suggests that the asymptotic behavior of log P({X > x})
is very similar to that of m
x
/ 2, and that the difference be-
tween them is asymptotically either a constant (as found in [7]
and [8] in a more restrictive setting) or a very slowly growing
function of x. Therefore, (4) provides more information on the
asymptotic behavior of P({X > x}) than (3), and suggests
that the simple approximation exp[m
x
/ 2] can be used to es-
timate P({X > x}), even for long-range dependent X
t
. In
Section IV, we will show that the improvement from (3) to (4)
can be critical for the accurate characterization (or estimation)
of the tail probability.
Here we should distinguish our work in this paper from
some results in the literature. All of the above discussion
(including the work in this paper) is about x-asymptotics
i.e., the asymptotic behavior of P({Q > x}), as the queue
length x increases. There has been recent work that focuses
on the asymptotic behavior of P({Q > x}) when the number
of sources, the queue length, and the service rate are all pro-
portionally sent to innity (e.g. [3], [16]). We classify these
studies as M-asymptotics, where M represents the number of
sources in the system. In particular, Montgomery and De Ve-
ciana [16] have signicantly strengthened the corresponding
0-7803-5420-6/99/$10.00 (c) 1999 IEEE
log-similarity relation in [3] using Bahadur-Rao asymptotics,
and obtained asymptotic bounds for the tail probability. How-
ever, note that M-asymptotics considers a limit in a differ-
ent direction from that in x-asymptotics. Therefore, results in
M-asymptotics cannot be extended to x-asymptotics (and vice
versa) unless very strong properties such as uniformity of con-
vergence hold (which is usually not the case). Hence, the re-
sults in this paper belong to a different category, from those in
M-asymptotics. However, as a side result of one of our propo-
sitions, we show in Section III, that we can further strengthen
the M-asymptotic result in [16] for Gaussian input processes.
As a nal note, due to space limitations, we do not provide
any proofs for the theoretical results in this paper. Interested
readers are referred to our technical report [6].
II. PRELIMINARIES
Throughout this paper we assume that = E{X
t
}/ t > 0
and that v
t
= Var{X
t
} is twice differentiable. Note that if
the input process
t
is a Gaussian process with stationary in-
crements (which, as mentioned before, is a good characteriza-
tion of the aggregate process to a high-speed multiplexer), and
the service process M
t
increases linearly with t at a xed rate
(i.e., a constant service rate which is typical in many high-
speed network models), then X
t
=
0
t
t is also a
Gaussian process with stationary increments.
We dene (t) := log v
t
and := lim
t
(t)/ log t (we
assume that the limit exists). It should be noted that from the
stationary increments property cannot be greater than 2. We
assume that [1, 2), which covers the majority of non-trivial
Gaussian processes with stationary increments. We next list a
few conditions on v
t
(and (t)) which will frequently be re-
ferred to throughout this paper.
lim
t
t
(t) = . (c1)
lim
t
t
2
(t) = . (c2)
v
t
t
St
< . (c4)
Conditions (c1) and (c2) are a direct result of the denition
of , (i.e., := lim
t
(t)/ log t) as long as LHospitals
rule can be applied. To elaborate, differentiating both numera-
tor ((t)) and denominator (log t), we get the left hand side of
(c1). Now if the limit in (c1) exists, it must equal . Similarly,
differentiating twice we get the negative of the left hand side of
(c2). Hence, if the limit in (c2) exists, it must equal . Con-
dition (c3) is closely related to the self-similarity of X
t
, i.e.,
if X
t
is (asymptotically) self-similar, then (c3) holds for some
> 1 (for more about the self-similarity and its origination,
see [14] and references therein). Also, a fairly general class of
long-range dependent X
t
satises (c3) for some > 1. Here it
should be noted that strictly second order self-similar Gaussian
processes, i.e., Fractal Brownian motion processes, constitute
but a very small subset of the long-range dependent processes
covered by our conditions with > 1. Condition (c4) is about
the behavior of v
t
around t = 0, and will be satised if v
t
de-
creases as fast as, or faster than t
as t 0. In particular, when
X
t
can be expressed as the integral of a stationary Gaussian
process, (c4) holds for any 2.
The parameter in our denition is directly related to the
well known Hurst (or self-similarity) parameter H by = 2H.
Also, the empirical estimate of has been popularly used to
observe self-similarity in various types of network trafc and
to calculate the corresponding Hurst parameters [2], [14].
We begin with an important property of the time-instant t
x
at
which v
t
_
(x +t)
2
attains its maximum value. More specif-
ically, we show that t
x
is a linear function of x.
Proposition 1 (Proposition 1 of [6]) Under hypothesis (c1),
t
x
x
x
(2 )
.
It should be noted that even if there are multiple indices at
which v
t
_
(x +t)
2
attains its maximum, Proposition 1 holds
for any choice of t
x
, among these indices. In fact, all the fol-
lowing results in this paper are independent of the choice of t
x
.
Also, as will be shown in the following section, under certain
conditions, t
x
becomes unique as x increases.
The next proposition is about the asymptotic behavior of
m
x
= (x +t
x
)
2
_
v
tx
= 1/
_
sup
t0
v
t
_
(x +t)
2
_
, and can
easily be derived using Proposition 1.
Proposition 2 (Proposition 2 of [6]) Under hypotheses (c1)
and (c3),
m
x
x
x
2
S
(2 )
2
.
For convenience, we dene a stochastic process {Y
(x)
t
: t
0} for each x > 0, as
Y
(x)
t
:=
m
x
(X
x
t
+xt)
x(t + 1)
.
From the denition of Y
(x)
t
, it directly follows that for all x >
0 and t 0,
X
t
> x if and only if Y
x
t
>
m
x
. (5)
Therefore, P({X > x}) is equal to P({Y
(x)
>
m
x
}).
This is important because we will study the supremum distri-
bution of X
t
through Y
(x)
t
. One can easily verify that Y
(x)
t
is
a centered (zero mean) Gaussian process, and its variance can
be obtained in terms of v
t
as
2
x,t
:= Var{Y
(x)
t
} =
m
x
v
x
t
x
2
(t + 1)
2
.
From the denition of m
x
, note that
2
x,t
attains its maximum
value of 1 at
t
x
:= t
x
/ x.
We next present the main results in the paper.
III. MAIN RESULTS
We rst demonstrate the importance of the time instant t
x
at
which the v
t
_
(x +t)
2
attains its maximum value.
0-7803-5420-6/99/$10.00 (c) 1999 IEEE
A. Importance of the Dominant Time Scale t
x
The rst theorem in this section is an improved version of
Theorem 4 in [8]. Informally, this theorem shows us how
the tail probability P({Y
(x)
>
m
x
}) will become concen-
trated around
t
x
= t
x
/x, as x increases.
Theorem 3 (Theorem 1 of [6]) Let > 0 and E
x
:=
_
t
x
x
2/2+
,
t
x
+x
2/2+
Ex
>
m
x
})
P({Y
(x)
>
m
x
})
= 1.
If we rewrite Theorem 3 in terms of conditional probability
as
lim
x
P({Y
(x)
Ex
>
m
x
|Y
(x)
>
m
x
}) = 1, (6)
then its implication becomes more evident. In other words,
(6) tells us that conditioned on the fact that the process Y
(x)
eventually exceeds a threshold (
m
x
), the probability that it
exceeds this threshold over a relatively small interval around t
x
will go to 1, as x goes to innity. Note that if we choose small
enough, the size of the interval E
x
around
t
x
decreases to 0 as
x increases. Hence, (6) tells us that the rarer the event becomes,
the more the event concentrates around the most likely time
t
x
.
Therefore, Theorem 3 can be interpreted as a rigorous veri-
cation of the statement rare events take place only in the most
probable way. Further, Theorem3 more directly explains why
the lower bound P({X
tx
> x}) = (
m
x
) so tightly bounds
the tail probability P({Y
(x)
>
m
x
}) = P({X > x}).
In our previous research [5], [7], we have numerically investi-
gated the accuracy of this lower bound as an approximation to
P({X > x}) under a more restrictive setting; i.e., when X
t
can be expressed as
X
t
=
_
t
0
d, (7)
where
t
is a stationary Gaussian process with negative mean
and absolutely integrable autocovariance. There, we have
found that the lower bound is fairly accurate and matches the
curve of P({X > x}) over a wide range of values of x. Fur-
ther, in [7], [8], we have shown that exp[m
x
/ 2] provides an
asymptotic upper bound to P({X > x}). Interestingly, this
asymptotic upper bound behaves (via empirical results) like a
global upper bound and approximates the tail probability as
accurately from above, as the lower bound does from below.
m
x
) and the asymptotic upper bound exp[m
x
/ 2] will
The empirical results make sense because exp[mx/ 2] is not very dif-
ferent from (
mx)
x
exp [mx/ 2]
2mx .
accurately approximate the tail probability, even for long-range
dependent X
t
. We will next provide an important asymptotic
property of P({X > x}) that supports this conjecture.
B. A Strong Asymptotic Result for P({X > x})
We begin this section by providing a theorem that shows
that the log of the tail probability log P({X > x}) (or equiv-
alently log P({Q > x}) diverges from m
x
/2 at most loga-
rithmically.
Theorem 4 (Theorem 2 of [6]) Under hypotheses (c1)(c4),
< liminf
x
1
log x
_
log P({X > x}) +
m
x
2
_
limsup
x
1
log x
_
log P({X > x}) +
m
x
2
_
< .
Note that (4) in the introduction is a compact form of Theo-
rem 4. The theorem suggests that m
x
/ 2 is a good estimate
of log P({X > x)} = log P({Y
(x)
>
m
x
}) in the sense
that the error could at most increase as (the order of) log x. In
other words,
P({X > x}) = exp
_
m
x
2
+r(x)
_
, (8)
where r(x) := log P({X > x}) + m
x
/ 2 O(log x).
We next relate our result with existing Large Deviation re-
sults. It should be observed that the leading term of m
x
, by it-
self, satises the Large Deviation relation (3), i.e., from Propo-
sition 2 and Theorem 4, we have
log P({X > x})
x
2
x
2
S
(2 )
2
. (9)
The right hand side of the above relation has been obtained
for the specic case of Fractal Brownian motion [17], and we
believe can also be obtained in greater generality by using the
results in [10]. If we dene R(x) := log P({X > x}) +
2
x
2
_
S
(2 )
2
, then the tail probability can also
be written as
P({X > x}) = exp
_
x
2
S
(2 )
2
+R(x)
_
. (10)
Further, it follows from (9) that R(x) o(x
2
)
where o(f(x)) denotes the set of functions g(x) such that
lim
x
|g(x) /f(x) | = 0. Since o(x
2
) is a much larger
set than O(log x), (4) characterizes the asymptotic tail behav-
ior in much more detail than (9), and therefore, signicantly
improves upon the resolution of (9). As will be illustrated
in the following section, this improvement can be critical for
accurately characterizing the asymptotic behavior of the tail
probability. For example, consider the two approximations
P({X > x}) exp
_
m
x
2
_
and (11)
P({X > x}) exp
_
x
2
S
(2 )
2
_
, (12)
naturally suggested by (8) and (10), respectively. From the
denitions of r(x) and R(x), exp[r(x)] and exp[R(x)] can
0-7803-5420-6/99/$10.00 (c) 1999 IEEE
be viewed as multiplicative factors that cause the error (i.e., its
deviation from 1 reects the inaccuracy of the approximation)
of the above approximations (11) and (12), respectively. Note
that the multiplicative factor of (11) can increase (or decrease)
at most as a power of x (e.g., on the order of x
3/2
or 1
_
x
2
),
while that of (12) can increase (or decrease) as an exponential
function of x (e.g., on the order of exp[
x] or exp
_
x
2/3
).
This indicates that we can signicantly reduce the error possi-
bility by using (11) instead of (12), for Gaussian X
t
.
An interesting remark is that the approximation for P({Q >
x}) based on the large deviation M-asymptotics result by
Botvich and Dufeld [3], results in the same expression as
exp[m
x
/ 2], when applied to Gaussian uid queues. Re-
member, from the introduction, that the M-asymptotics result
in [16] improved upon the result in [3] (from log-similarity
to nearly similarity), and an approximation equivalent to the
lower bound (
m
x
) was suggested based on these stronger
asymptotics. Although both the lower bound and the approxi-
mation (11) provide accurate estimates for the tail probability,
in [7], it has been shown that when the tail is asymptotically
exponential (i.e., the input is not long-range dependent), the
(logarithmic) error the lower bound diverges to , while that
of (11) is bounded. This tells us that the approximationthat sat-
ises only the weaker asymptotics in M-asymptotics [3], may
perform better in the x-asymptotics sense (and vice versa). As
mentioned in Section I, this is because x-asymptotics and M-
asymptotics consider asymptotic properties of P({Q > x}) in
different limiting regimes.
As mentioned earlier, the uniqueness of t
x
(or
t
x
) is not
a major issue in this paper because all the results are valid
for any choice of t
x
when there are multiple indices where
v
t
_
(x +t)
2
attains its maximum. However, from a practical
viewpoint, it may be important to know whether t
x
is unique
or not, and how easily we can nd t
x
and compute the value
of m
x
. For example, when exp[m
x
/ 2] has to be computed
to approximate P({X > x}), its accuracy and computation
time will be determined by how fast and accurately we can cal-
culate m
x
. Since both t
x
and m
x
cannot generally be obtained
in a simple closed form, search algorithms are likely to be used
for the computation of t
x
and m
x
. In this case, the following
proposition guarantees the accurate and fast computation of t
x
and m
x
, for large x.
Proposition 5 (Proposition 3 of [6]) Under hypotheses (c1)
and (c2), for all sufciently large x, log v
t
_
(x +t)
2
is
strictly concave on [x/2, ( +
_
/2)x/(2 )], and
there is a unique index t
x
where v
t
_
(x +t)
2
attains its max-
imum.
Proposition 5 tells us that when x is large, t
x
and m
x
can be
computed by performing a simple local search algorithm start-
ing at x/(2 ). Although this proposition is valid only
for large enough x, according to our numerical studies, local
search algorithms usually nd t
x
and m
x
accurately within a
small number of iterations, even for fairly small values of x.
This is because v
t
_
(x +t)
2
is usually of a distinctly uni-
modal shape even for small x.
It is also worth to note that Proposition 5 can strengthen the
M-asymptotics result in [16] to a similarity relation. In [16,
Theorem A.1], it has been shown for general uid queues that
1 liminf
M
c
M
P({Q
M
> Mx})
limsup
M
c
M
P({Q
M
> Mx}) K,
where Q
M
represent the queue length distribution of a queue
that serves M identical input processes at service rate M (for
the denition of c
M
, refer to the paper). When the input pro-
cesses are stationary Gaussian, it turns out that K is the num-
ber of points t where
2
Mx,t
of the queue attains its maximum.
Therefore, Proposition 5 implies that when x is sufciently
large, we have K = 1 and the above relation can be rewrit-
ten in a stronger form as
P({Q
M
> Mx})
M
1
c
M
.
In the next section we study the asymptotic properties of
m
x
in more detail, and the effect of its secondary terms on the
asymptotic behavior of P({X > x}).
IV. THE IMPACT OF Var{X
t
} ON THE ASYMPTOTIC
BEHAVIOR OF m
x
AND P({X > x})
In this section we will introduce our third theorem, which
relates the asymptotic behavior of Var{X
t
} to that of m
x
and
P({X > x}). We begin with a simple example of Fractal
Brownian motion processes, a well known and studied set of
self-similar processes [17].
A. Fractal Brownian Motion Process
The standard (normalized) Fractal Brownian motion process
{B
(H)
t
: t 0} with Hurst parameter H [1 /2, 1) is a
centered Gaussian process with stationary increments that pos-
sesses the following properties [17]:
(a) B
(H)
0
= 0,
(b) Var{B
(H)
t
} = t
2H
,
(c) B
(H)
t
is sample path continuous.
We rst study the supremumdistribution of X
t
:= SB
(H)
t
t
which is often called Fractal Brownian motion with negative
linear drift.
From the above properties of Fractal Brownian motion pro-
cesses, one can easily verify that X
t
satises all conditions
(c1)(c4) with = 2H. Also, in this case we can compute m
x
explicitly by
m
x
=
4
x
2
S
(2 )
2
. (13)
Therefore, for Fractal Brownian motion processes with nega-
tive linear drift, (9) can be strengthened (by substituting (13)
into (4)) to
log P({X > x}) +
2
x
2
S
(2 )
2
O(log x). (14)
In other words, for Fractal Brownian motion with negative lin-
ear drift, the approximation based on the Large Deviation tech-
niques can in fact be better supported by Theorem4. Of course,
0-7803-5420-6/99/$10.00 (c) 1999 IEEE
Simulation
e e
m
x
x
S
=
2
2
2
2
2
T
a
i
l
P
r
o
b
a
b
i
l
i
t
y
:
P
X
x
>
Supremum Value: x
Fig. 2. The tail probability P({X > x}) for Xt = B
(0.8)
t
2t computed
by simulations and (12). Since Xt is a Fractal Brownian motion process
with drift, (11) and (12) result in the same approximation.
note that in this case (11) and (12) results in the same approx-
imation. Due to this fact, and our numerical studies [7] on
the accuracy of the approximation (11), we expect that (12)
will be an accurate approximation over a wide range of x.
As an example, see Fig. 2 where we plot P({X > x}) for
X
t
= B
(0.8)
t
2t. Since the exact tail probability cannot be
computed, its estimate has been obtained through simulations.
In particular, in order to improve the reliability of the simu-
lation results, the Importance Sampling technique described
in [12] has been employed. To show the accuracy of the es-
timate, 99% condence intervals of the simulation results have
been computed using the method of batch mean [4]. How-
ever, the simulation results are so accurate that the condence
intervals are virtually invisible in the gure. Since the simu-
lation has been performed in discrete-time, the tail probability
estimated via simulations is actually the probability that X
t
ex-
ceeds x at any integer t. Hence, the simulation results provides
an accurate lower bound for the tail probability P({X > x}).
Nevertheless, one can see that the approximation (12) accu-
rately captures the tail probability over the entire range of x
shown in the gure.
x
2
S
(2 )
2
O(log x) (15)
Therefore, if (15) holds, the approximation (12) can be better
supported by (14), just as strongly as the approximation (11) is
supported by (4).
We next provide our third theorem which shows, with more
precision than Proposition 2, how the asymptotic behavior of
Similar observations have been made in [17]. In that paper, the author
used the lower bound
x
2
S
(2 )
2
to approximate
the tail probability. However, since the Gaussian tail function (w) was again
approximated by exp
w
2
1.
Under hypotheses (c1)(c4), if there exists a positive constant
C such that
t|h
x
2
S
(2 )
2
x
2
x
2
S
(2 )
2
h(t
x
).
Corollary 7 (Corollary to Theorem 6) Under hypotheses (c1)
(c4), if there exists a positive constant C that satises (16) and
if h(t) = v
t
_
St
1 O(t
2
log t), then (15) holds.
The additional condition over (c1)(c4) that Theorem 6 re-
quires, is not a very restrictive condition since as will be illus-
trated in the following section, we can usually nd a constant
C > 0 that satises (16). Therefore, Theorem 6 not only re-
sults in a sufcient condition (Corollary 7) for (15) to hold, but
also tells us in considerable generality when (15) will not hold.
Let u
t
include all auxiliary terms of v
t
(i.e., terms other than
the leading term St
+u
t
.
Further assume that u
t
t
ct
t
ct
x
2
S
(2 )
2
x
2c
2
x
2+2
S
2
2
(2 )
2+2
.
(17)
In other words, Theorem 6 tells us that asymptotically, the
more slowly growing the auxiliary term u
t
is, the more
m
x
/ 2 behaves like 2
x
2
_
S
(2 )
2
. In this
sense, the Fractal Brownian motion process with negative lin-
ear drift that we have just studied, is an extreme case, where
u
t
= 0 for all t and (14) trivially holds. Hence, we next con-
sider cases when u
t
is not identically equal to 0, and show that
(14) does not hold in general.
B. Other Gaussian Processes
From (17), if (2 2, ), then r(x) R(x) increases
as a power of x, and since r(x) O(log x), it follows that
R(x)
x
2c
2
x
2+2
S
2
2
(2 )
2+2
. (18)
Therefore, if (2 2, ), then (14) will not hold. Through
a simple example, we now show that can, in fact, be greater
than 2 2.
Assume that 1 < < 3/ 2, and consider a stationary Gaus-
sian process
t
with mean and autocovariance given by
E{
t
} = and (19)
Cov{
t
,
t+
} =
S( 1)
2(|| + 1)
2
. (20)
0-7803-5420-6/99/$10.00 (c) 1999 IEEE
Simulation
e
x
S
2
2
2
2
e
m
x
2
T
a
i
l
P
r
o
b
a
b
i
l
i
t
y
:
P
X
x
>
Supremum Value: x
Fig. 3. Two approximations based on (4) and (9) for P({X > x}), when Xt
is dened by (21), and t is a stationary Gaussian process with E{t} =
1 and Cov{t,
t+
} = 5(|| + 1)
3/4
8.
If we dene
X
t
:=
_
t
0
d for t 0, (21)
then one can easily verify that X
t
is a Gaussian process with
stationary increments that satises (c1)(c4), and that
E{X
t
} = t and
Var{X
t
} = S((t + 1)
1) St. (22)
From (22), it can easily be veried that u
t
= v
t
St
St and that th
(t)
t
(1 )h(t) (a sufcient condition
for (16) to hold for some C). Therefore, we actually have =
1 > 2 2 in this case, and it follows from (18), that
R(x)
x
2
21
x
32
S
22
(2 )
32
.
Further, from (10) note that the approximation (12) will
involve an error (for large x) roughly on the order of
exp
_
2
21
x
32
_
S
22
(2 )
32
n=1
(n)
d.
Also, from (1), the tail P({Q > x}) of the steady state
queue length distribution is the same as P({X > x}), where
{X
t
: t 0} is dened by X
t
:=
_
0
t
L
n=1
(n)
d. Note
that the variance of X
t
can be expressed in terms of the auto-
covariance C
n
() := Cov{
(n)
t
,
(n)
t+
} of L input processes
as
v
t
=
L
n=1
2
_
t
0
(t )C
n
()d. (23)
In other words, v
t
is composed of L terms, each of which is
determined by the autocovariance of the corresponding input
process. If we assume that 2
_
t
0
(t )C
n
()d
t
S
n
t
n
for some S
n
> 0 and
n
[1, 2), one can easily see that
v
t
t
S
1
t
1
, where
1
= max{
1
,
2
, . . . ,
L
} and
S
1
=
{n:n=
1}
S
n
. Therefore, if the values of
n
are not iden-
tical, the leading term
S
1
t
1
of v
t
will capture only the terms
in (23) that increase on the order of t
1
, and hence u
t
is likely
to increase on the order of t
2
, where
2
is the second largest
value among {
1
,
2
, . . . ,
L
}. If
2
is greater than 2
1
2
in this case, then from (18), R(x) will grow on the order of
x
2+
22
1
, and the approximation (12) will be poor for large
x.
V. NON-GAUSSIAN INPUT PROCESSES
As mentioned in Section I, Gaussian uid queueing models
for high-speed networks have been motivated by the large num-
ber of network applications served by a single network link. In
other words, since the aggregate trafc at a high-speed mul-
tiplexer will be the superposition of a large number of packet
(or information) streams, each of which is from different ap-
plications, the Central Limit Theorem suggests that the aggre-
gate trafc should be well characterized by a Gaussian pro-
cess. Although modeling the trafc in a high-speed network as
Gaussian processes is a intuitive and natural approach, the er-
ror that this Gaussian trafc modeling may involve, is usually
difcult to analyze. For this reason, the validity of the Gaus-
sian trafc modeling has been numerically investigated in [5],
[7]. Through this numerical study, we found that the behavior
of a uid queue fed by several hundreds of independent in-
put processes can typically be captured by the Gaussian model
in the sense that the original aggregate trafc and its Gaus-
sian model result in very similar queueing behavior. As a re-
sult, the approximation (11), which is originally for Gaussian
uid queues, is also accurate when applied to the analysis of
high-speed networks. Next, through a numerical example, we
demonstrate the accuracy of the approximation (11) for long-
range dependent non-Gaussian inputs.
0-7803-5420-6/99/$10.00 (c) 1999 IEEE
Simulation
e
m
x
2
69 sources
66 sources
T
a
i
l
P
r
o
b
a
b
i
l
i
t
y
:
(
)
P
Q
>
x
Queue Length: x
Fig. 4. The approximation (11) for 66 and 69 MPEG video sources served by
a 45 Mbps link compared to the simulation results.
In [2], video applications have been found to be typical
sources of long-range dependent network trafc. In the fol-
lowing example, we compare the approximation (11) and the
simulation results of the tail probability at a multiplexer serv-
ing video applications with a 45 Mbps link. For the experi-
ment, we use the real trace of an MPEG-encoded action movie
(007 series) used in [13].
In this paper, the trace has been found to exhibit long-range dependence.
For more details about the trace, refer to [13] or [18].
and/or self-similar behavior. As illustrated in the previous ex-
ample, under this heterogeneity, the Large Deviation result (9)
may not be precise enough to capture the queueing behavior of
the network trafc, and an approximation like (12) should be
used with caution.
In contrast, (4) considerably improves the resolution of (9),
and naturally leads to the approximation (11), which will not
rapidly diverge from P({X > x}). Therefore, we hope that
the results in this paper will be important in better understand-
ing the behavior of the supremumdistribution of Gaussian pro-
cesses with stationary increments, and in analyzing the queue
length distribution for heterogeneous types of network trafc.
REFERENCES
[1] R. G. Addie and M. Zukerman, An Approximation for Performance
Evaluation of Stationary Single Server Queues, IEEE Transactions on
Communications, vol. 42, pp. 31503160, Dec. 1994.
[2] J. Beran, R. Sherman, M. S. Taqqu, and W. Willinger, Long-range de-
pendence in variable-bit-rate video trafc, IEEE Transactions on Com-
munications, vol. 43, pp. 15661579, Feb.April 1995.
[3] D. D. Botvich and N. G. Dufeld, Large deviations, the shape of the loss
curve, and economies of scale in large multiplexers, Queueing Systems,
vol. 20, pp. 293320, 1995.
[4] P. Bratley, B. L. Fox, and L. E. Schrage, A Guide to Simulation. New
York: Springer-Verlag, second ed., 1987.
[5] J. Choe and N. B. Shroff, A New Method to Determine the Queue
Length Distribution at an ATM Multiplexer, in Proceedings of IEEE
INFOCOM, (Kobe, Japan), pp. 550557, 1997.
[6] J. Choe and N. B. Shroff, Supremum Distribution of Gaussian Pro-
cesses and Queueing Analysis including Long-Range Dependence and
Self-Similarity, tech. rep., Purdue University, West Lafayette, IN, 1997.
submitted to Stochastic Models.
[7] J. Choe and N. B. Shroff, A Central Limit Theorem Based Approach
for Analyzing Queue Behavior in High-Speed Networks, IEEE/ACM
Transactions on Networking, vol. 6, pp. 659671, Oct. 1998.
[8] J. Choe and N. B. Shroff, On the supremum distribution of integrated
stationary Gaussian processes with negative linear drift, Advances in
Applied Probability, March 1999. to appear.
[9] A. Dembo and O. Zeitouni, Large Deviations Techniques and Applica-
tions. Boston: Jones and Bartlett, 1993.
[10] N. G. Dufeld and N. OConnell, Large deviations and overow proba-
bilities for the general single server queue, with application, Proc. Cam-
bridge Philos. Soc., vol. 118, pp. 363374, 1995.
[11] P. W. Glynn and W. Whitt, Logarithmic asymptotics for steady-state tail
probabilities in a single-server queue, Journal of Applied Probability,
pp. 131155, 1994.
[12] C. Huang, M. Devetsikiotis, I. Lambadaris, and A. R. Kaye, Fast Sim-
ulation for Self-Similar Trafc in ATM Networks, in Proceedings of
IEEE INFOCOM, pp. 438444, 1995.
[13] E. W. Knightly, Second Moment Resource Allocation in Multi-Service
Networks, in Proceedings of ACM SIGMETRICS, (Seattle, WA), 1997.
[14] W. E. Leland, M. Taqqu, W. Willinger, and D. V. Wilson, On the
Self-Similar Nature of Ethernet Trafc (Extended Version), IEEE/ACM
Transactions on Networking, vol. 2, pp. 115, Feb. 1994.
[15] R. M. Loynes, The Stability of a Queue with Non-independent Inter-
arrival and Service Times, Proc. Cambridge Philos. Soc., vol. 58,
pp. 497520, 1962.
[16] M. Montgomery and G. De Veciana, On the Relevance of Time Scales
in Performance Oriented Trafc Characterization, in Proceedings of
IEEE INFOCOM, (San Francisco, CA), pp. 513520, 1996.
[17] I. Norros, On the Use of Fractal Brownian Motion in the Theory of
Connectionless Networks, IEEE Journal on Selected Areas in Commu-
nications, vol. 13, pp. 953962, Aug. 1995.
[18] O. Rose, Statistical properties of MPEG video trafc and their impact
on trafc modeling in ATM systems, in Proceedings of the 20th Con-
ference on Local Computer Networks, (Minneapolis, MN), pp. 397406,
Oct. 1995.
[19] N. B. Shroff and M. Schwartz, Improved Loss Calculations at an ATM
Multiplexer, IEEE/ACM Transactions on Networking, vol. 6, pp. 411
422, Aug. 1998.
0-7803-5420-6/99/$10.00 (c) 1999 IEEE