Вы находитесь на странице: 1из 18

ELSEVIER Performance Evaluation 22 (1995) 239-256

An approximate model for performance evaluation of real-time


multimedia communication systems *
Stanley S. Wang * and John A. Silvester
Department of Electrical Engineering Systems, University of Southern California, Los Angeles, CA 90089-2562, USA

Abstract

A Markov Modulated Poisson Process (MMPP) is a doubly stochastic Poisson process which has recently received
a lot of attention due to its ability to model a highly correlated arrival stream while retaining analytical tractability.
In this paper, the cell arrival process from integrated voice and data sources is approximated by a two-state MMPP
with batch arrivals (BMMPP). We propose a new matching technique which leads to more accurate performance
prediction than other recent work for both delay estimation and evaluation of tail probabilities. While using BMMPP
in the model does not noticeably increase the computational complexity over MMPP, it allows better representation
of traffic sources with significantly higher burstiness. Another major contribution of the proposed model is its
simplicity which makes the model suitable for real-time traffic control. An extended version of the model for
integrated video, voice and data sources is also given with numerical examples again showing that the accuracy is
quite satisfactory.

Keywords: Performance evaluation; Markov modulated process; Multimedia communications

1. Introduction

The communication industry is currently in the process of migrating its infrastructure to


Asynchronous Transfer Mode (ATM) based Broadband Integrated Services Digital Networks
(B-ISDN). The goals of B-ISDN, i.e., a wide range of new services and effective utilization of
network resources, present a great challenge in its traffic control. This is due to the fact that
the traffic sources for the future B-ISDN are expected to be varied in their statistical
characteristics and Quality of Service (QoS) requirements. For examples, data traffic typically

” The equipment used in this research was procured as a combination of gift from Hewlett Packard and from
funds from the National Science Foundation under Grant NO. CDA-9121961.
* Corresponding author. Current address: Computer Science Department, California State University, San
Marcos, CA 92096-0001, USA.

0166-5316/95/$09.50 0 1995 Elsevier Science B.V. All rights reserved


SSDI 0166-5316(94)00057-3
240 S.S. Wang, J.A. Siluester / Performance Evaluation 22 (1995) 239-256

has a very strict loss requirement, while voice traffic is usually delay sensitive. Hence, in
modeling these traffic sources, it is necessary to use different models for different types of
traffic, e.g., data, voice and video. Using different models for different types of traffic, however,
increases the complexity of the traffic model. For real-time traffic control, where processing
time is critical, a simple approximate traffic model allowing fast yet accurate performance
prediction is needed.
A Markov Modulated Poisson Process (MMPP) with batch arrivals (denoted by BMMPP in
this paper), a special case of the Neuts’ versatile Markovian process (the N-process [17], which is
called the Batch Markov Arrival Process, BMAP, by Lucantoni in [12]), is a doubly stochastic
Poisson process with batch arrivals where the batch arrival rate is determined by the state of the
underlying finite-state Markov chain. BMMPP is suitable for modeling traffic of the future
B-ISDN since it allows representation of highly bursty and correlated arrivals yet retains
analytical tractability in the analysis of their queueing behavior [12,17]. Some results related to
Markov modulated queueing systems are available in the literature. Knessl and Mathkowsky
study the stationary distribution of the unfinished work for a two-state MMPP/G/l queue
[lo]. Prabhu and Zhu complete a detailed analysis, including the waiting time, the idle time and
the busy period, of an MMPP/G/l queue [18]. Recently, Zhu, in his work [26], includes new
results for an MMPP/M/l queue with bulk arrivals. In [20], Ramaswami details the analysis of
an N/G/l queue including the arrival process, the stationary queue length at departures, and
the waiting time. Most of these results have been used extensively in traffic and performance
modeling (see for examples: [8,X,22-24]). In this work, we specialize the results presented in
[12] and use a two-state BMMPP to study the performance of ATM multiplexers loaded with
various types of traffic.
There are two main factors that contribute to loss in a queueing system with bursty sources:
short-term and long-term overloads. Short-term overloads are caused by a burst of arrivals that
cause loss for a small buffer. Long-term overloads are caused by several sources being at their
peak arrival rates at the same time. This may persist for several tenths of a second which
corresponds to several thousand of cell times for a typical B-ISDN channel and cannot be
resolved by additional buffer space (at least for a reasonable amount of buffer space). In [l] and
[24], for example, it has been shown that the cell loss due to long-term overload has very limited
reduction over a certain buffer size. Hence, there is limited value in further increasing the
buffer size. We assume that the systems under study are well dimensioned (meaning that cell
losses due to the short-term randomness of the arrival process occur rarely and that cell losses
are mainly due to long-term overloads) and concentrate on the delay performance of the
systems.
We model both the superposition of voice sources and the superposition of video sources as
separate birth-death processes (these models are discussed in the next section). The cell arrival
process from the aggregation of data sources is modeled as a Poisson process with batch
arrivals. We develop a low complexity parameter matching procedure and propose an approxi-
mation for the integrated voice and data arrival process using a two-state BMMPP. An
integrated services ATM multiplexer is then modelled as a two-state BMMPP/D/l queue.
The performance of the proposed approximation is verified by comparison to a similar model
recently proposed by Baiocchi et al. [l] and simulation results. The results show that our model
leads to quite accurate prediction of the system delay and a reasonably good evaluation of the
S.S. Wang, J.A. Silvester/ Performance Evaluation 22 (1995) 239-256 241

Fig. 1. An ON-OFF process.

tail probabilities. We then extend the approximation to model the integration of video, voice
and data sources also by a two-state BMMPP with a different parameter matching procedure.
The numerical results again show that the resulting approximation is quite satisfactory. In
practice, this proposed approximation can be used in real-time traffic control because of its
simplicity.
The paper is organized as follows: more discussion of the traffic models for voice and video
sources is found in Section 2. The approximation technique is discussed in Section 3.
Simulation and numerical results are presented in Section 4. Finally, Section 5 provides some
conclusions.

2. Models for voice and video traffic

Numerous studies that characterize packetized voice sources, e.g., [1,4,5,8,9,11,14,21,22,25],


have been made. Among these studies, Brady’s ON-OFFprocess [5] has been widely adopted as
the model for the arrival process corresponding to a single voice source. In the ON-OFF
process model, the voice source alternates between exponentially distributed ON periods and
exponentially distributed OFF periods. The ON periods correspond to talkspurts while the
OFF periods represent silence periods. Packets are created with a constant interarrival time
during the ON periods and no packets are generated during the OFF periods. The transition
rates, from-ON-to-OFF (p) and from-OFF-to-ON (a), are determined by the expected dura-
tions of the talkspurts and the silence periods (see Fig. 1). We can represent the superposition
of N such ON-OFF processes by a finite-state birth-death process with the state representing
the number of voice sources in talkspurt (in the ON-state) as shown in Fig. 2.
We assume that the video sources are of the video-telephony type (showing a person talking
in front of the camera without sudden movements) and use a continuous-time Murkov chain to
model the superposition of several video sources, as suggested by Maglaris et al. [13]. In this
model, the output bit rate from a codec is assumed to fall into a set of discrete bit rate levels
evenly spaced between zero and the peak rate. Transitions are assumed to occur between
adjacent levels only, according to a birth-death Markov chain. The transition rates are
determined by matching the model’s parameters with statistical characteristics of the traffic, as
described in [13]. Fig. 3 shows the model with A4 + 1 bit rate levels, 0, 77, 217,. . . , Mq. Note

Fig. 2. Model for superposition of N voice sources.


242 S.S. Wang, LA. Silvester / Performance Evaluation 22 (1995) 239-256

Fig. 3. Model for video sources with a total of M + 1 bit rate levels.

that this model can be decomposed into A4 independent ON-OFF type mini-sources each of
which has parameter set a, 6 and 7.
The difficulty in generating useful analytical results from a queueing system with an arrival
process based on the birth-death processes defined above has forced researchers to look for
approximations. Two main approaches have been proposed. The first approach uses a fluid
flow approximation [6,15,22], which overlooks the short-term variations of the arrivals and has a
limited range applicability. The second approach carefully matches the parameters of the
process to a simpler one, e.g., the two-state MMPPs used in [1,8,15], and the renewal process
used in [15]. This approach still has high computational complexity which limits its use for
real-time systems applications. We therefore modify the approach of [l] reducing the complex-
ity and also improving the accuracy.

3. The approximation technique

The cell arrival process corresponding to superposed voice sources is approximated by a


two-state MMPP as in [l] but a different procedure for parameter matching is proposed. We
use a similar approach to include video sources into our model by observing that integration of
the models for voice and video sources results in a two-dimensional Markov chain. Data packets
are assumed to form a Poisson process, i.e., data cells form a Poisson process with batch
arrivals. The arrivals from the integrated video, voice and data sources are then modeled as a
two-state MMPP with batch arrivals.

3.1. Using two-state MMPP as an approximation for voice sources

Heffes and Lucantoni, in their work [8], use the following parameters of the two-state MMPP
to match with the underlying Markov chain model of the superimposed voice sources:
(i) the expected arrival rate;
(ii) the variance-to-mean ratio of the number of arrivals in some time interval;
(iii) the long-term variance-to-mean ratio of the number of arrivals;
(iv) the third moment of the number of arrivals in some time interval.
Upon observing the strong role of the overload period (defined to be when the number of
active sources exceeds the capacity of the system) in determining the performance of the
multiplexer, Nagarajan et al. in their matching procedure [15], replace step (iv) in [8] by “the
variance of the number of arrivals in some time interval giving that the system is in overload
state”, and achieve an improvement in predicting packet loss for a finite-buffered system. In [l],
Baiocchi et al. attack the problem from a different angle. They use a theorem proved in [16]
(Theorem 2.3.1, p. 62) and present an “asymptotic matching procedure” which leads to more
S.S. Wang, J.A. S&ester /Performance Evaluation 22 (1995) 239-256 243

accurate results than those of [S] and [22] (see the reference for details). Our approach is
similar but has lower complexity and achieves higher accuracy.

3.2. The matching technique

Consider a multiplexer loaded with N independent voice sources each of which is modeled
by an ON-OFF process with the following parameters:
(i) w, the constant arrival rate in the ON-state;
(ii) (Y,the transition rate from the OFF-state to the ON-state;
(iii) p, the transition rate from the ON-state to the OFF-state.
Let C denote the capacity (in cells per unit of time excluding the average capacity consumed
by data sources, if any) of the multiplexer; L = [C/o] be the maximum number of active voice
sources which the multiplexer can support; Ti be the expected time until the superposition
process of the iV ON-OFF processes visits state i - 1 for the first time starting from state i; and
rk defined by

~k=($+Jk(-&)N-k, OlklN

be the steady state probability that k out of the N voice sources are in the ON-state.
To find q, let us first observe that it satisfies the following recurrence relation:
1 (N-i)a
?= (N_i)a+ip + (N_i)cu+ip(~+l+T,)7 lli<N-1

with the boundary condition, TN = l/(Np). Note that, in (21, the first term on the right-hand
side gives the expected sojourn time in state i. The first part of the second term on the
right-hand side specifies the probability that the process will make a transition to state i + 1
given that it is currently in state i. It can be easily proved, by successive substitution, for
example, that TN_i satisfies the following equation:
i!
_ai-k

TN-i= i
k=O
(N-k:)! ’
(N-i- l)!p’+lPk

Thus, TN_i has the following explicit closed-form expression:

TN-i = i:I) iO(r)(i)i-k, O<i<N-1. (3)


W i
Using (3), we can find the expected time until the system load drops below the system capacity,
TL+ 1, once the load exceeds the system limit, L.
Our matching technique works as follows: referring to Fig. 4, the states of the superposition
arrival process are divided into two disjoint subsets, the overload states, {L + 1, L + 2,. . . , NJ,
244 S.S. Wang, J.A. Silvester / Performance Evaluation 22 (1995) 239-256

~:::::::::~~:::::~. . . --;TJ<::::::;~ . . .D

Fig. 4. State mapping between the arrival process and the two-state MMPP for superposition of N voice sources.

and the underload states, (0, 1, 2,. . . , L}. These two sets of states are mapped into the state-II
(the overload-state) and the state-I (the underload-state) of a two-state MMPP respectively. In
Fig. 4, y1 and y2 are the transition rates and h, and A, are the cell arrival rates for state-I and
state-II correspondingly. We then propose the following approach to match the parameters
(note that, (ii - iv) are effectively the same as those used in [l]):

II, = c rTTk
;
k=O
N
where 17, = C rk ;
k=L+l

is the expected number of active calls.

As one might expect, the arrival stream from the superposed process is burstier than that
from the corresponding two-state MMPP especially when the two-state MMPP is in its
overload-state (when there is a relatively larger number of voice sources in talkspurt). This can
be understood by observing the “smoothing” effect introduced by the reduced number of states
in the approximate model, because when the two-state MMPP is in either one of its states the
arrival process is just an ordinary Poisson process rather than each active source emitting cells
at its peak rate. Since, on the average, a longer queue is expected by a new arrival for a burstier
arrival process due to a shorter interarrival time within the same burst, the approximation is
expected to perform worse as the system load increases, i.e, the number of active sources and
the burstiness of the arrival process increase. The same problem is also experienced in the
asymptotic matching procedure proposed in [l].

3.3. Improve the approximation

In order to improve the accuracy of our model, let us observe that if we overestimate TL+l,
we may, in fact, improve the model’s accuracy. This is because the arrival rate in the
X.7. Wang, J.A. Silvester / Performance Evaluation 22 (1995) 239-256 245

overload-state is higher than that in the underload-state (see step (ii) and (iii) of the matching
procedure). Although increasing the average overload time also causes an increase in the
average underload time (see step (iv) of the matching procedure), staying in the overload-state
longer each visit would mean a higher burstiness (for the same overall arrival rate). Therefore,
overestimating the average overload time, TL +1, will increase the burstiness of the two-state
MMPP. Note also that (for the same number of voice calls) the number of states which are
mapped into the overload-state of the two-state MMPP increases, i.e., L decreases, as the
channel capacity decreases. This corresponds to the two-state MMPP staying a longer time in
the overload-state on the average. In addition, the longer the average overload time is, the
more severe the underestimation of the burstiness is. Thus, a higher overestimation for TL +I is
needed to offset the increasingly severe underestimation of the burstiness.
One approach that overestimates TL+ I is to replace the expression in (3) by its upper-bound.
We found that the following upper-bound for (31, which replaces the summation in (3) by its
largest term times the number of terms in the summation, leads to very good results:

(4)

By applying Stirling’s fmmda [71 to (41, we have

i-k
i+l NN
O~;~.“(fi)
< -Cl kk(N-k)N-k
NP

i+l
=
NP

xO$;;iO(fl)exp N log N-(N-k)log(N-k)-k log k+(i-k)log

i+l

N-k N-k k k
x OF;;iO(fl)exp - Floghi’- Elogz I (5)

The problem now reduces to determining the value of k that provides the largest exponent in
(5). Let us r ep 1ace k/N by x and rewrite the exponent of (51 as

N (X - 1) log(1 -.x) --x log x + (i-X) log(;)].


i
246 S.S. Wang, J.A. Silvester / Performance Evaluation 22 (1995) 239-256

After taking the derivative of (6) and equating to zero, we have

log(l-x)-log x-log ; =o. (7)


i i
From (7), it can be easily verified that (6) is maximized at k = N/3/(a + p>. Thus, we conclude
that if i < N/3/(a + p), then k = i provides the largest term in (4); otherwise, k = kN/3/((r + p>]
should be chosen. In the following theorem, we prove that for the systems of interest
N - L - 1 < Np/(a + p> always holds. Thus, (4) leads to a very simple expression for i = N - L
- 1 and can be used in the matching procedure instead of using (3).

Theorem 1. For any stable system, i.e., utilization p = (Ncuo)/[C(cy + p)] < 1, N-L - 1<
N[P/(a + p)] holds.

Proof. Assume the theorem is not true, i.e., for some (Y,p, and N

NP
N-L-l>----
ff+p

=++I.,_(l-%)N

NCtW
- (a+@C >
yyl-(l-$)
*p>l
which contradicts the assumption. Therefore, the theorem is true. q

Theorem 1 implies that for the systems of interest, i < Np/((r + p) always holds for
i = N - L - 1. Therefore, the upper-bound that we are interested in, denoted TL+ is given by I,

Note that, in the extreme case, i.e., when L + 1 = N, f,,, is equal to TN. Furthermore, for the
same number of voice _calls, as the value of L + 1 decreases, i.e., the ch_annel capacity
decreases, the value of (T, +1 - TL +1>increases. This is precisely how we want TL +1 to behave,
as pointed out before. As we will demonstrate by numerical examples (see Figs. 6-8, when
applied to our model, this simple expression leads to a much better result. Another point worth
noting is its simplicity which is very important especially for real-time traffic control.
S.S. Wang, J.A. Silvester /Performance Evaluation 22 (1995) 239-256 247

Fig. 5. Transition rate diagram and state mapping for integrated video and voice jources.

We can now refine our matching procedure using the result of (8) to

(L + OP.
(i) y2=-+--=
L-t1
N-L ’

(ii)- (as before).

3.4. Include video traffic to the model

To motivate the discussion of our extension to include video sources into the model, let us
observe that the expected time that the system will stay in the overload state depends on two
major factors: for the same number of traffic sources, the heavier the system load is, the longer
each visit to the overload state is; and, the quicker the system can reduce its current load, the
faster it will leave the overload state. Note that, in (S), (N - L)/( L + 1) is the ratio of the
number of overload states to the number of underload states, which gives a measure of how
heavy the system load is. The other term, l//3, is the expected time until an active source
becomes idle, which is an indication of how soon the current system load will drop.
To include video traffic into the model, we combine the models for voice and video sources
which results in a two-dimensional Markov chain, as seen in Fig. 5. The states represent the cell
arrival rate level for video sources (M + 1 levels in all) and the number of active voice sources
248 S.S. Wang, J.A. Silvester/ Performance Evaluation 22 (1995) 239-256

(N voice sources total). The steady-state distribution of this two-dimensional Markov chain,
which can be easily derived from its local balance equations and global balance equations, is
given by

We say a state (i, j) is overloaded if it satisfies the following condition:


(qi + wj) > C. PO)
Otherwise, we say that it is underloaded. Recall that 77 is the step size of the cell arrival rate
levels for video sources and w is the cell arrival rate for an active voice source. Let S,, S, be
the sets and 8,, 8, be the number of the underload states and overload states respectively in
Fig. 5; and ai be the number of voice sources which the system can support given that the video
sources are currently in cell arrival rate level i. Note that ai must satisfy the following
conditions: Us= [(C - qi)/o] and - 1 5 a, IN (where oi = - 1 counts the situation where the
video sources in cell arrival rate level i alone overload the system).
From the definition of ai, it can be easily seen that
M M
o,= c (N-a,) and 19,= c (ui + 1). (11)
i=O i=O

By adopting a similar idea to that used for voice sources, we then propose to use
1
&!S.-
(12)
8, P+b
as an estimation of the average time the system stays in the overload-state. Referring to Fig. 5,
we then propose the following matching procedure for the two-state MMPP as an approxima-
tion of the traffic due to integrated voice and video sources:

(i) y2=$=>(P+b);
*

(ii) A, = c (qi + oj) , where IIfi = c rXy;


(i, j)ESU (x, Y)ES,

(iii) A, = c (qi + wj) , where Ilo = c n._,;


6, YkS,

NCX Ma
where & = - and 41~= -
a+p a+b’

As before, the first term of f provides an indication as to how heavy the system load is and
the second term measures how quickly the current system load will drop. Note that the latter is
the mean of a random variable that is the minimum of the following two exponentially
distributed random variables: (a) the time that a voice source stays in the ON-state with mean
S.S. Wang, J.A. Siloester / Performance Evaluation 22 (1995) 239-256 249

of l/p and (b) the time a video mini-source stays in its ON-state with mean of l/6. Note also
that step (ii) calculates the average cell arrival rate when the process is in an underload
situation. Similarly, the third step finds the expected cell arrival rate when the process is in an
overload period. The last step matches the average arrival rate of the two-state MMPP with
that of the actual cell arrival process.

3.5. Add data traffic to the model

We have thus reduced the system description to a two-state MMPP to represent the arrival
process from a number of video and voice sources. To include data sources into the model, let
us assume that the integration of data sources generate data packets according to a Poisson
process with rate p. The packet length has an arbitrary probability mass function, Pk = Pr {data
packet is comprised of k cells}. This is the same as saying that the data-cell arrival process is a
batch-arrival Poisson process with a random batch size. If we combine the data sources with the
two-state MMPP, we have a two-state MMPP with batch arrivals. The transition rates, y1 and
yZ, are determined by the matching procedure. The arrival rates are h, + Al.and h, + p for the
underload-state and the overload-state respectively. The probability that a batch has a size of k
cells is
*i
PkL +8k-1-- i= 1, 2; k= 1, 2, 3 ,... (13)
p +hi p +&

where

1, if k= 1,
6 k-l =
0, otherwise.

4. Numerical results and discussions

In this section, we use the algorithm described in the Appendix to calculate the average
system time and use a specialized algorithm also presented in the Appendix to find the tail
probabilities of a BMMPP/D/l queue (with infinite buffers). We verify the proposed approxi-
mation by comparing numerical results to the results from [l] as well as to simulation.
Numerical examples for voice and data integration are presented first followed by numerical
and simulation results for video, voice and data integration. Some discussions are included at
the end of this section.

4.1. Numerical results for voice and data integration

We model a voice/data ATM multiplexer as a continuous-time two-state BMMPP/D/l


queue. The algorithm (which makes use of the results presented in [12]) presented in Appendix
is applied to find the expected system time. In these examples, each voice call is characterized
by o = l/6 cells/ms (this corresponding to a 64 Kbps PCM coding with speech activity
250 S.S. Wang, J.A. Siivester / Performance Evaluation 22 (1995) 239-256

Fig. 6. Expected system time V.S.channel utilization for 20 voice calls and 20 data calls.

detector and a standard 48-octet payload size), (Y= 1.538 and p = 2.778 (according to the
conclusions drawn by [4]). The aggregated data packet arrival rate is assumed to be 2ON,/3
packets per second, where iVd (an user-defined parameter) is the number of data calls. We
assume that the packet size is geometrically distributed with an average of 5 cells per packet. In
all figures presented in this sub-section, BMMPP-1 is the model suggested by [l] extended to
include data sources. BMMPP-2 is the model proposed by this work using Eq. (8) and
BMMPP-3 is the result using Eq. (3). For Figs. 6 and 7, we fix the system load (by having a
constant number of voice and data calls) and plot average system time versus channel
utilization. 20 voice calls plus 20 data calls and 200 voice calls plus 200 data calls are assumed
for Figs. 6 and 7 correspondingly. For Fig. 8, which shows the relationship between average
system time and channel capacity, a constant system utilization of 0.9 and a fixed ratio of the
number of voice and data calls of 3 : 5 are assumed. To further demonstrate the superiority of
the proposed model (in comparison to the one proposed by 1111,in Figs. 9 and 10 we use the
algorithm presented in Appendix to plot the tail probabilities of a system with 20 voice calls
and 50 data calls. Note that for the analytical result 5 presented in Figs. 9 and 10 (corresponding
to a system utilization of 0.7 and 0.9 respectively), we assume an error tolerance of 10m2’
instead of solving the system with infinite buffers (see the Appendix for details).

Fig. 7. Expected system time V.S.channel utilization for 200 voice calls and 200 data calls.
S.S. Wang, J.A. Siluester/Performance Evaluation 22 (1995) 239-256 251

Channel GapaciIy (h4bp.s.)

Fig. 8. Expected system time V.S.channel capacity for p = 0.9.

1 co 1
so 100 160 200

Buffer Position

Fig. 9. Tail probabilities for 20 voice calls and 50 data calls with p = 0.7.

4.2. Numerical results for video, voice and data integration

We use the same voice and data sources as in the previous examples and use the same set of
parameters for the video sources as used by [3] and [13], i.e., video sources are characterized by:

‘c [ .. .
B -....
I_
3 -.
s .%.
-..
J
a
2c 0

Fig. 10. Tail probabilities for 20 voice calls and 50 data calls with p = 0.9.
252 S.S. Wang, J.A. Silvester / Performance Evaluation 22 (1995) 239-256

Number of Video Sources

Fig. 11. Expected system time V.S. number of video sources for p = 0.8.

an average bit rate of 3.9 Mbps, a peak bit rate of 10.58 Mbps, a standard deviation of the bit
rate of 1.73 Mbps and a parameter for the autocorrelation function of 3.9. The total number of
discrete cell arrival rate levels for video sources is assumed to be 16 times the number of video
sources. In Fig. 11, we assume 100 voice sources and 100 data sources in the background and
plot the relation between average system time and the number of video sources under a system
utilization of 0.8 (varying the channel capacity). Our model is shown as curve BMMPP-2 and
for comparison purposes we plot a curve BMMPP-1 using the minimum of l/p and l/b rather
then l/(p + b) as the measure of how fast the current system load will drop. In Fig. 12, we
keep the ratio of the number of video, voice and data calls fixed at 1: 20 : 20 and show average
system time versus channel capacity with different (constant) system utilizations (i.e., by varying
the total load).

5. Discussions and conclusions

We first point out that for the model presented in [l], solving an (N - L)-dimentional eigen
system is required in order to find the parameters for the MMPP. This requires a significant
amount of computation effort especially when N - L is large, i.e., when the load of the system

Channel Capacity (Mbps)

Fig. 12. Expected system time V.S. channel capacity for different system utilizations.
S.S. Wang, J.A. Silvester / Perfomance Evaluation 22 (1995) 239-256 253

is heavy. Yet, as demonstrated by Figs. 6 and 7, the model failed to make a good delay
estimation of the modeled system under these situations. Whereas in our model, the computa-
tional requirement is insignificant and is independent of the system size. More importantly, our
model provides better delay performance predictions (agrees with the simulation better) for any
given system load; and, the improvement is more noticeable for a higher system load. For video,
voice and data integration, although we provide no comparison to other models (since no
similar models exist), simulation results show that the approximation is very good even under
heavy system load (Figs. 11 and 12).
Several advantages of our model can be identified. First, as discussed above, less computa-
tional overhead is involved. Second and more importantly, it provides better performance
predictions. Third, a real-time traffic control algorithm can be developed. Note that if the system
is in an overload situation, the queue size is expected to build up until the system returns to an
underload state. Hence the expected length of the overload periods, 1/y2, can be used as an
indicator for the seriousness of the overload, once the system is overloaded. Using our results,
l/y, can be computed very quickly for any given number of video, voice and data calls and
used as a criterion for traffic control.
In summary, we study the performance of an ATM multiplexer loaded with video, voice and
data traffic. The actual arrival process is approximated by a two-state Markov modulated
Poisson process with batch arrivals; and the multiplexer is modeled as a two-state BMMPP/D/l
queue. The modeling technique developed here leads to very accurate results and has very low
computational requirements. Hence it is useful for modeling real-time traffic. Some studies
have been done which prove the effectiveness of this approach for estimating loss probability of
systems with limited buffer space (see [24]). Due to the discrete-time nature of ATM networks,
we have also developed a similar discrete-time model in [23] which also has good accuracy.

A. Appendix

In this appendix, we include procedures for calculating the expected system delay and the
queue length distribution of a BMMPP/D/l queue. These procedures, which are presented in
a form specifically for a two-state BMMPP/D/l queue, can be easily extended to a procedure
for a BMMPP/D/l queue with any number of states.
Throughout this appendix, boldfaced uppercase characters and arrowheaded characters are
used to represent matrixes and vectors correspondingly. We use &(z> to represent the
Z-transform of a discrete probability distribution function f,, e.g.,S%(Z) means the Z-transform
of D,. We also adopt the following additional notation:
l A, and yi are the arrival rate and transition rate of state i of the two-state BMMPP
respectively.
l <= (qi, qz) = (y2/(y1 + y2), yI/(yI + yz)) is the steady state probability of the underly-
ing Markov chain of the two-state BMMPP.
l h = A,q, + h,q, is the average arrival rate of the two-state BMMPP.
l gi(k) is the probability of a bulk size being k given that the system is in phase i.
l S(t) = u(t - T) is the service time distribution, where u( *) is the unit step function and T
is the constant service time.
254 S.S. Wang, J.A. Silvester / Performance Evaluation 22 (I 995) 239-256

l (AJij, the (i, j) component of matrix A,, yt = 0, 1, 2,. . . ,1 I i, j I 2, is the probability


that given the system is not empty and the arrival process is in phase i at the beginning of
a service period, there are n arrivals during the service period and the phase of the
two-state BMMPP is j at the end of the service period.
l (BJij, the (i, j) component of matrix B,,n = 0,1, 2,. . . , 1 I i, j I 2, is the probability
that given the system is left empty by a cell departure and the arrival process is in phase i
right after the departure, there is a total of II + 1 arrivals until the next cell departure
instant and the phase of the two-state BMMPP is j at the end of this period.
l xii, 1 I i I 2, 0 ~j I K- 1, is the steady state joint probability of the phase being i and
the queue length being j at cell departure instants (right after each departure).
l 2) = xii + xZi, 0 I i I K - 1, is the steady state probability of the queue length being i.

. X=(x‘,, ?p..,T&l 1, where Xi = (xii, xZi>, 0 I i I K - 1, is the steady state joint proba-
bility distribution of the queue length and the phase of the two-state BMMPP at cell
departure instants.
l Y, which has a similar structure as X, is the steady state joint probability distribution of
the queue length (including server) and the phase of the two-state BMMPP at any time
instant.

A.1. Average system delay for two-state BMMPP/D /l queue

The algorithm based on [12] for calculating the expected system time, w, of a two-state
BMMPP/D/l queueing system is given as follows:
(i) Construct D, = [-~;+y~)_$+,,ZJ, D, = [h$‘l(k)hPg,(kJ V k 2 1, and D = CzEOD, =
r-77 YI 1.
LY2
(ii) Comp$e d= E~=,kD,Z, where e= (:I;
(iii) Solve G = 10”e o(G)s dx = C~=,(/~ e -ex(8x)k/k! dS(x))(I +&(G)/e)k, where L%(G)
= C;=-Jl,Gk, I = [by] and 8 = maxJ --D(Glii) (note that: starting with G = e$, the
algorithm suggested by [12] can be used recursively to solve G);
(iv) Compute g’ using ZG = g‘ and 6 = (1 - p>g’ where p is the system utilization;
(VI Compute y1 = -T&‘(l) and Q_= T2(&‘(1) +6”(l));
(vi) Compute u1 = V,c and c2 = V,e;
(vii) Compute w, = l/2(1 - p)[2p + 26 - ijTIXZi+ D)-‘G, + $T21;
(viii) Compute i?= -w,~-~+(~-_ql/,)(~‘q+D)-‘;
(ix) The average system time is then given by w = -($),a‘)-‘i&&?+ T&“(1)~/2~~

A.2. Queue length distribution for two-state BMMPP /D / 1 queue

We summarize an algorithm to compute the queue length distribution of a two-state


BMMPP/D/l/K queue. In practice, for K = 03 (i.e., a BMMPP/D/l queue with infinite
buffers), the A,‘s and B,'sare calculated up to a certain number where the truncation error is
small enough to be ignored. According to our experience, for an error tolerance of 10e2’
(which is assumed to be the case in this work for solving the tail probabilities of a BMMPP/D/l
queue), around 190 A,‘s and B,'sare required to be calculated.
S.S. Wang, J.A. Silvester/Performance Evaluation 22 (199.5) 239-256 255

One can specialize the algorithm presented by Lucantoni [12] to that of BMMPP/D/l/K_
queues for computing A,‘s and B,'s. Once these matrices are known, we can easily see that X
is the invariant probability vector of the following stochastic matrix:

Bo 4 4 ... BK-2 iI
i=K-1
Bi

4 Al A2 ... AK-2 C Ai
i=K-1
m
R=
0 4 AI ... AK-3 C Ai
i=K-2

...
00
0 0 0 ... A0 CAi
i=l

Thus, we can solve 2 directly using the facts that i?R=i? and CF=&,‘xli+xli = 1. Specifically,

j&z. (R)_’

where b = (0, 0 ,...,O,1) and 15 = R -I with all elements of its last column replaced by 1. As
given in [12], the Z-transform for the queue length distribution at any time is

7&z) =h(l -z)~@&+)[zz-~(2)]-‘.


By comparing the coefficients of ti in the above equation, yi can then readily obtained.

References

[l] A. Baiocchi et al., Loss performance analysis of an ATM multiplexer loaded with high-speed on-off processes,
IEEE J. Select. Areas Comm. 9 (1991) 388-393.
[2] C. Blondia, The N/G/l finite capacity queue, Stoch. Models 5 (2) (1989) 273-294.
[3] C. Blondia and 0. Casales, Performance analysis of statistical multiplexing of VBR sources, Proc. INFOCOM’92,
pp. 6C.2.1-6C.2.11.
[4] P.T. Brady, A statistical analysis of on-off pattern in 16 conversations, Bell Systems Tech. .I. (January 1968)
73-91.
151 P.T. Brady, A model for generating on-off speech patterns in two-way conversation, Bell Systems Tech. J.
(September 1969) 2445-2472.
[6] J.N. Daigle and J.D. Langford, Models for analysis of packet voice communications systems, IEEE J. Select.
Areas Comm. 4 (1986) 847-855.
[7] R.L. Graham, Concrete Mathematics: A Foundation for Computer Science, Addison-Wesley, Menlo Park, Calif.
(1989).
[8] H. Heffes and D.M. Lucantoni, A Markov modulated characterization of packetized voice and data traffic and
related statistical multiplexer performance, IEEE .I. Select. Areas Comm. 4 (1986) 856-868.
[9] I. Ide, Superposition of interrupted Poisson processes and its application to packetized voice multiplexers, Proc.
ZTC ‘88, pp. 3.1B2.
[lo] C. Knessl and B.J. Matkowsky, A Markov modulated M/G/l queue I: stationary distribution, Queueing
Systems 1 (1987) 355-374.
256 S.S. Wang,J.A. Silvester/ Performance Evaluation22 (1995) 239-256

[ll] S.-Q. Li, A general solution technique for discrete queueing analysis of multimedia traffic on ATM, ZEEE
Trans. Comm. 39 (1991) 1115-1132.
[12] D. Lucantoni, New results on the single server queue with a batch Markov arrival process, Stoch. Models 7 (1)
(1991) l-46.
[13] B. Maglaris et al., Performance models of statistical multiplexing in packet video communications, IEEE Trans.
Comm. 36 (1988) 834-844.
[14] B. Mukherjee, Integrated voice-data communications over high-speed fiber optic networks, Computer (February
1991) 49-58.
[15] R. Nagarajan et al., Approximation techniques for computing packet loss in finite-buffered voice multiplexers,
IEEE J. Select. Areas Comm. 9 (1991) 368-377.
[16] M.F. Neuts, Matrix-Geometric Solutions in Stochastic Models: An Algorithmic Approach, Johns Hopkins Univer-
sity Press, Baltimore, Md. (1981).
[17] M.F. Neuts, A versatile Markovian point process, J. Appl. Prob. 16 (1979) 764-779.
[18] N.U. Prabhu and Y. Zhu, Markov modulated queueing systems, Queueing Systems 5 (1989) 215-246.
[19] G. Ramamurthy and B. Sengupta, Delay analysis of a packet voice multiplexer by the CD, /D/l queue, IEEE
Trans. Comm. 39 (1991) 1107-1114.
[20] V. Ramaswami, The N/G/l queue and its detailed analysis, Adu. Appl. Prob. 12 (1980) 222-261.
[21] K. Sriram and W. Whitt, Characterizing superposition arrival processes in packet multiplexers for voice and
data, IEEE J. Select. Areas Comm. 4 (1986) 833-846.
[22] R.C.F. Tucker, Accurate method for analysis of a packet-speech multiplexer with limited delay, IEEE Trans.
Comm. 36 (1988) 479-483.
[23] S.S. Wang and J.A. Silvester, A discrete-time performance model for integrated service ATM multiplexers,
Proc. GLOBECOM’93, pp. 757-761.
[24] S.S. Wang and J.A. Silvester, Estimate the loss and tail probabilities for multimedia communications systems,
Proc. GLOBECOM’94, pp. 908-912.
[25] C. Yuan and J.A. Silvester, Queueing analysis of delay constrained voice traffic in a packet switching system,
IEEE J. Select Areas Comm. 7 (1989) 729-738.
[26] Y. Zhu, A Markov modulated M/M/l queue with group arrival, Queueing Systems 8 (1991) 255-264.