AN EFFICIENT APPROACH FOR SPEEDING UP

SIMULATION OF WIRELESS NETWORKS

 y
Frank H.P. Fitzek, Edjair Mota, Enno Ewers, Adam Wolisz

Te hni al University of Berlin, Germany

z
K. Pawlikowski

University of Canterbury, Christ hur h, New Zealand

Keywords : Sto hasti simulation, automated distributed simulation,

wireless networks, performan e evaluation

ABSTRACT tage of using this approa h in performan e eval-

uation studies su h omplex and highly dynami
There have been an in reasing demand for wireless sto hasti systems as wireless ommuni ation net-
ommuni ation te hnologies and therefore there works. Both the issue of speeding up the simula-
exists a ne essity of designing eÆ ient tools for tion, and the importan e of produ ing results with
enhan ing the omputational eort of their perfor- an appropriate (small) statisti al error, will be ad-
man e analysis. Wireless ommuni ation systems dressed.
are hara terized by highly variable parameters as-
so iated with the wireless hannel itself as well as
with hanges in hannel performan e dependent 1 INTRODUCTION
on the number of a tive users. Be ause of that sim-
ulation studies of su h networks may be very time- There have been an in reasing demand for wireless
onsuming. A simple and ee tive way of speeding ommuni ation te hnologies and therefore there
up su h simulations an lead through employment exists a ne essity of designing eÆ ient tools for
of distributed pro essing power of modern om- enhan ing the omputational eort of their per-
puter networks. In this paper we report results forman e modeling and analysis. Stru tural om-
of a performan e study of a wireless network ob- plexity of modern tele ommuni ation networks
tained by applying one spe ial s enario of sto has- auses that in many situations omputer simu-
ti simulation, known as Multiple Repli ations In lation is the only way of investigation of these
Parallel, or MRIP, in whi h multiple pro essors highly dynami sto hasti systems. For exam-
operate as individual simulation engines, generat- ple, in the problem investigated in this paper, a
ing output data that are submitted to a global wireless ommuni ation network, operating under
analyzer for on-line data analysis. The global ana- Code Division Multiple A ess (CDMA), expe-
lyzer is responsible for stopping (distributed) sim- rien es highly variable onditions, aused by ran-
ulation when the statisti al error of results rea hes dom properties of ommuni ation hannels, mu-
the required level. MRIP is an attra tive s enario tual interferen e of signals generated by multiple
for speeding up omputationally intensive sto has- a tive users, et . Simulation provides a powerful
ti simulation sin e it an be fully automated. tool to prototype su h networks during the design,
The goal of this paper is to show up the advan- as well as during the development and operational
 This Work has been supported partially by a grant phases.
from the BMBF (German Ministry for S ien e and Te h- Sto hasti nature of the pro esses o urring in
nology) within the Priority Program ATM mobil. tele ommuni ation networks means that one needs
y Supported by Conselho Na ional de Desenvolvimento
to produ e the nal results with a known (small)
Cient o e Te nologi o, from Brazil, under grant 29.0003-
94.6. statisti al error. Additionally, in all simulation
z This work was supported by the Alexander von Hum- studies aimed at assessing long-run performan e of
boldt Foundation networks, the output data are highly orrelated, so
they have to be analyzed using spe ial statisti al Be ause of the initial omplete un ertainty
te hniques. Development of a urate methods of about the length of steady-state simulation needed
statisti al analysis of orrelated simulation output for obtaining results with a satisfa tory (small)
data has attra ted a onsiderable s ienti interest statisti al error, and having suspe ted that very
and eort. For a thorough treatment of this and long simulation times would be required, we di-
related problems see [11℄. Unfortunately, straight- re ted our attention on an implementation of
forward simulation of omplex networks an re- MRIP s enario of distributed sto hasti simula-
quire prohibitively long simulation times despite tion in AKAROA-2 [17℄, a pa kage oering auto-
of in reasing power of modern omputers. It's not mati ontrol of the nal error of simulation re-
unusual for a simulation experiment to take days sults, as well as automated parallelization of (or-
or weeks, for yielding results with an a eptable dinary) simulation programs on multiple pro es-
level of statisti al error. In this situation, a on- sors of a lo al area network. For additional speed-
siderable interest has been fo used on parallel and ing up of our simulation experiments, we linked
distributed simulation of tele ommuni ation net- AKAROA-2 with PTOLEMY [16℄, a pa kage reg-
works. Speeding up exe ution of simulation is a ularly used in the Tele ommuni ation Networks
hallenging resear h problem whi h has attra ted Group at the Te hni al University of Berlin, Ger-
a onsiderable s ienti interest and eort. Mu h many, for onstru ting simulation models. The
of the resear h has been dire ted toward a s enario advantages of using AKAROA-2 in performan e
where many pro essors ooperate in exe uting of evaluation studies of wireless networks are dis-
a single repli ation of simulation. A somewhat ussed in Se tion 3. We look there at the issue of
under-valued s enario uses many pro essors that speedup of simulation under MRIP, as well as at
are engaged into running their own repli ations of the importan e of drawing on lusions about the
the simulated system and ooperate with entral quality of performan e of tele ommuni ation net-
analyzers (one entral analyzer for ea h perfor- works on the basis of results with small statisti al
man e measure analyzed), that are responsible for errors.
observing the stopping riteria of the simulation.
We refer to this approa h as Multiple Repli ations
In Parallel (MRIP) (see [12℄ for a dis ussion of its 2 SIMULATION
properties in the ontext of steady-state simula-
tion), whereas the former is generally known as SCENARIO AND MODEL
Single Repli ation In Parallel (SRIP). The analyzed s enario of a wireless network on-
Despite of its simpli ity, the MRIP approa h sists of a spe i number of Wireless Terminals
seems to resolve satisfa torily the two main prob- (WTs) operating within one wireless ell. All WTs
lems of sto hasti simulation of tele ommuni ation ommuni ate with one entral Base Station (BS),
networks. Su h a on lusion an be drawn on the whi h overage denes the ell boundaries (see g-
basis of performan e evaluation study reported in ure 1). We assume a CDMA based mobile om-
this paper. We have investigated performan e of muni ation system with a number of odes mu h
a wireless network operated under CDMA, using higher than the number of a tive WTs. The wire-
the Binary Phase Shift Keying (BPSK). The an- less link is onsidered to be unreliable with a vary-
alyzed network onsists of one wireless ell and a ing Bit Error Probability (BEP). The value of
varying number of Wireless Terminals (WTs) and the BEP depends on the number of used hannels
one entral Base Station (BS). The wireless links k . For the hosen s enario we assume that trans-
are onsidered to be unreliable, and subje ted to mission may be ae ted by an Additive White
Additive White Gaussian Noise (AWGN). The Gaussian Noise (AWGN) and the Binary Phase
purpose of this study was to investigate the long- Shift Keying (BPSK) is used. The basi bit rate
run quality of servi e experien ed by users of su h of a CDMA hannel is 64 kbit/s.
a network, measured by the jitter of su essfully Using the simplest ARQ me hanism (Send and
transmitted data segments (as a fun tion of the Wait ), like it is dis ussed in [1℄ and suggested
segment length, and the number of a tive mobile by the re ent wireless LAN standards [6, 7℄, ea h
users) and by the segment loss probability. The erroneous pa ket is retransmitted. The following
simulated model is dis ussed in more detail in se - stored pa kets have to wait until the pa ket has
tion 2. been transmitted su essfully. So, during that time
 hannels the overall noise level will in rease, re-
11111
00000
00000
11111
sulting also in a lower ee tive bit rate for all WTs.
MS 1
00000
11111
00000
11111
00000
11111
00000
11111
Therefore me hanisms are needed whi h prote t
00000
11111
the system against performan e degradation due
MS 6 to the SMPT approa h. For example, a base sta-
tion, whi h has full knowledge about the dimen-
1
0
0
1 1
0
1 1
0 0
1
0 MS 2
sion of the system, might assign only a limited
MS 7 1
0 1
0
number of hannels to ea h mobile. Alternatively,
11111
00000
00000
11111
a mobile might stop using ex essive number of
00000
11111
00000
11111
00000
11111
00000
11111
hannels if the throughput does not improve. This
MS 5 00000
11111 MS 3
leads to the assumption that an optimal number
of hannels an be found, whi h has low in uen e
MS 9
on other wireless mobile terminals and assures sta-
bility of bandwidth. After bad hannel onditions
MS 4 MS 8
during start  hange , we ompensate the bit rate
degradation ! = (Bgood Bbad )
(start  hange )
by using multiple CDMA hannels. In order to re -
Figure 1: Wireless S enario: Nine Terminals Com- ognize the hanges of the link quality an informa-
muni ating With One Base Station tion feedba k is needed. In general a CDMA radio
front-end an support several hannels in paral-
lel. It is obvious that there exists a platform de-
pendent number of hannels whi h an be used in
the ee tive bit rate de reases from Bgood to Bbad parallel by one mobile. Depending on the system
and simultaneously the jitter for single MAC pa k- properties this maximum number an even equal
ets, as well as for segments at higher layers of one, although a number of hannels up to 8 seems
network's proto ols, in reases. We assume that a to be realisti .
resulting bit rate Bbad is not a eptable for the The main analyti al problem is to investigate
required throughput spe ied by the QoS param- the in uen e of SMPT on the QoS parameters in
eters. Further the in reased jitter is not a eptable terms of jitter and segment losses, as well as these
for the appli ation. Simultaneous MAC Pa ket ee ts with those ones hara terizing ordinary se-
Transmission (SMPT) is a method that over omes quential transmission.
the problem of instable bandwidth and high de-
lay variation by utilizing parallel hannels (see
[3℄). The main ideas are as follows. As long as no X
stop
B hannel = Bgood
( hange start ) (1)
error o urs the pa kets are transmitted sequen-  hange
tially using one CDMA hannel. When a pa ket
is orrupted at start , the sender will re ognize As mentioned, the simulations have been per-
the hange in hannel state by the missing a - formed using the Ptolemy simulation tool [16℄ and
knowledgment. Under the assumption that errors AKAROA-2 [17℄. It required linking AKAROA-
on wireless hannels are orrelated (bursty), the 2 with the Ptolemy interfa e akstars (see gure
sender will probe out the hannel by sending prob- 2). We formed a ommuni ation system with 3; 6;
ing pa kets (see [4℄) at a rate adjusted to the han- and 9 WTs and one BS. The hannel between the
nel's onditions. At  hange , when a probing pa ket WT and the BS was modeled with a multilayered
is su essfully a knowledged, the sender suddenly Markov hain, onsidering two hannel states (bad
takes the advantage of using parallel hannels to and good ) and the impa t of used hannels on the
orre t the ae ted jitter. The sender will pro- BEP. The main parts of the simulation model are
eed to used multiple hannels up to stop , deter- the proto ol implementations of the WT and BS.
mined by equation 1. The joint bit rate Bjoint;k Ea h WT generates a stream of transport units
using k CDMA hannels of bit rate Bi , is gener- (like UDP segments and therefore alled segments )
ally smaller than the sum of the bit rates of all with a spe i load, and passes these segments to
hannels. This is be ause the higher number of the DLC layer via the network layer, where ea h
used hannels results in a higher bit error probabil- of them is divided into a group of DLC pa kets. To
ity [2, 5℄. With an in reasing number of allo ated ea h pa ket a header with length  is added. This
header  is used to identify DLC pa kets in the
right order and to assign the DLC pa kets to the
appropriate segments and equip them in the means
for error dete tion. The frame, whi h is omposed
by one DLC pa ket and the header  is alled a
Data Link Control Pa ket Data Unit (DPDU).
All DPDUs are stored in a queue with a xed
buer apa ity LQueue within the DLC layer and
will be sent with dierent ARQ based transmission
methods over the wireless link. There is no error
orre tion s heme assumed, so pa kets with one
or more bit errors will not be de oded su essfully
on the re eiver side and will be onsidered lost.
If a segment an not be transmitted within given
maximum delay and/or a given bounded jitter, the
sender side MAC entity will stop transmitting this
segment and pro eeds with the next segment. The
dis arded segment will be ounted as lost. The re-
eiver side MAC re ognizes this loss by dete ting
the in reased segment number in the header of the
next re eived pa ket. The load generation module
generates pa kets with onstant bit rate, allowing Figure 2: Ptolemy Simulation S enario with Nine
variable as well as xed segment sizes at a spe i Wireless Terminals and the Akaroa Star
load level. We negle t the fading ee ts and as-
sume an optimal power ontrol within the WTs.
Nevertheless, using the wireless hannel, ea h WT
will in uen e other WTs by an in reased ba k- for sequential determination of the length of the
ground noise. To get a feeling how SMPT will in- initial transient stage, based on a sequential ver-
uen e the QoS oered by a wireless link the re- sion of S hruben's test for testing stationarity of
sulted jitter will be investigated. time series; see [14℄ and [11℄. On the other hand,
output data in steady-state were analyzed by ap-
plying the method of Overlapping Bat h Means
3 SIMULATION OUTPUT (OBM), proposed by Meketon and S hmeiser [9℄.
Our preliminary investigations of the OBM have
DATA ANALYSIS AND shown that this method is very robust on ern-
FINAL RESULTS ing its underlying assumptions (e.g. independen e
among bat hes), and yields onden e intervals
As mentioned, all simulation experiments reported ontaining the theoreti al values with probability
in this paper were exe uted using AKAROA-2. lose to the nominal onden e level, espe ially in
This means that the a ura y of all estimates was the ase of highly dynami sto hasti pro esses.
automati ally assessed sequentially during simu- An analysis of the performan e of this method, as
lation, and the simulation was stopped when sta- applied to queuing networks, an be found in [10℄.
tisti al errors of the results rea hed an assumed As one an expe t, in all onsidered s enarios
a ura y. That was measured by the relative sta- of the investigated wireless network the length of
tisti al error, dened as the ratio of the half-width simulation was in reasing with the number of WTs
of the onden e interval and the point estimate, and the assumed onden e level of results. Fig-
at a given onden e level. To avoid using the re- ure 3 shows the speedup of simulation exe uted
sults obtained by too short simulation runs we under AKAROA-2 on 10 pro essors, for 3 dier-
ran ea h experiment three times, using dierent ent simulated s enarios. The speedup is measured
sequen es of pseudo random numbers, and a - here as the ratio of the exe ution time of simula-
epting the results produ ed by the longest run tion on one pro essor T(1), and the exe ution time
only[15℄. Condu ting steady-state simulation, we of simulation on 10 pro essors T(10). One an see
have used a standard pro edure of AKAROA-2 an important property of MRIP : it performs even
 11111
00000
WT P=1 P = 10
10.69 00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
3 I 0.51/5103/243 0.51/7290/972
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
II 0.52/5187/247 0.51/6072/765
000000
111111 00000
11111
00000
11111 III 0.50/4662/259 0.51/7554/1000
T1 / T10

7.43 000000
111111
000000
111111 00000
11111
00000
11111
000000
111111
6.05 11111
00000
00000
11111
00000
11111
000000
111111
000000
111111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
6 I 0.51/8928/248 0.54/4464/744
00000
11111
00000
11111
00000
11111
00000
11111
000000
111111
000000
111111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
II 0.52/8877/269 0.51/3228/538
3.75 11111
00000
000000
111111
00000
11111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
000000
111111
000000
111111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111 III 0.50/9216/256 0.53/4572/762
000000
111111
00000
11111
000000
111111
000000
111111
00000
11111
00000
11111 11111
00000
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
9 I 0.53/10458/249 0.55/9777/486
2.74
00000
11111
000000
111111
000000
111111
00000
11111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
000000
111111
000000
111111
00000
11111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
3 000000
111111
00000
11111
000000
111111
WTs111111 00000
11111
6 WTs 11111
000000 9 WTs
00000 00000
11111
3 6 WTs 11111
WTs111111
000000
000000
111111
00000
11111
00000
00000
11111
9 WTs
00000
11111 II 0.53/10038/239 0.53/13623/518
00000
11111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111 000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
000000
111111
000000
111111
00000
11111
00000
11111 000000
111111
000000
111111 00000
11111
00000
11111 III 0.52/10248/244 0.53/12820/494
0.95 0.99
Confidence Level
Table 2: Simulation Results for 3, 6 and 9 WTs
Figure 3: Redu tion in Time by Applying MRIP with (Conden e Level of 99% and Relative Error
with Ten Pro essors for Dierent S enarios (3 up of 5%)
to 9 Wireless Terminals)

better if one applies it to longer simulations. the initial transient phase of the simulation, i.e.
Although we have used an almost homogeneous the number of observations dis arded before the
set of pro essors, the above results an be to some steady state had begun.
extent ina urate, as they an vary with the sys- One an lear observe that under MRIP, dis-
tem load of a given Lo al Area Network (LAN). A tributing simulation on more pro essors is mu h
better way of studying this phenomenon would be more eÆ ient, as one may need fewer observa-
to analyze an average number of observations re- tions for stopping simulation (with the same level
quired for stopping the simulation when it rea hes of statisti al error) than in the ase of fewer pro-
an a eptable level of statisti al error. essors. Dierent (short) lengths of the transient
phases justify the appli ation of the sequential sta-
WT P=1 P = 10 tionarity tests (su h as the test implemented in
3 I 0.51/3645/243 0.51/2916/486 AKAROA-2), instead of dis arding an arbitrary
II 0.52/3705/247 0.51/3108/518 (xed) number of observations. Dis arding too few
III 0.50/3885/259 0.52/2964/494 initial observations an lead to statisti ally biased
6 I 0.51/4464/248 0.54/2976/496 nal results [11℄, while dis arding too many obser-
II 0.51/4035/269 0.51/1614/269 vations would unne essarily lengthen the simula-
III 0.50/4608/256 0.51/3072/512 tion.
At the end, let us look at the SMPT results. Fig-
9 I 0.53/2988/249 0.53/4494/749
ure 4 and 5 show the histogram of the jitter of su -
II 0.53/3585/239 0.54/5736/956
III 0.52/3660/244 0.52/4410/735 essfully transmitted network layer segments as a
fun tion of the segment length (measured in MAC
Table 1: Simulation Results for 3, 6 and 9 WTs pa kets). Obviously the SMPT 1 method provides
(Conden e Level of 95% and Relative Error of better results in the sense of jitter. Future appli a-
5%) tions will need stable and bounded jitter in ombi-
nation with low segment losses. The resulted seg-
ment loss probabilities are shown in table 3 and
Table 1 and Table 2 depi t the results hara - re e ts the fa t that with SMPT we an expe t
terizing the performan e of the analyzed wireless performan e improvements not only the sense of
network in the three s enarios (with 3, 6 and 9 the jitter but also in the sense of segment losses.
WTs), obtained when using 1 or 10 pro essors, Therefore we laim that SMPT is able to support
and assuming 95% and 99% onden e levels of future multimedia appli ations more eÆ ient than
results, respe tively. For ea h wireless terminal in other proposals.
ea h s enario, the rst value stands for the jit- The results presented in Figure 4 and 5 were
ter estimate, the se ond value stands for the sam- obtained with the relative error of 1% (at the
ple size required for a hieving the relative error of 1 Only three odes were allowed to be used simultane-
5%, and the third value stands for the length of ously.
 0.025
0.016

0.014

0.02

0.012

0.01 0.015
0.02 0.03

Probability
Probability

0.015
0.008 0.02

0.01 25 25
0.01 0.01
0.005 0.006
20 20

s]
s]

et
et
0

ck
0
ck

15

pa
15
pa

0.004 0
0

AC
AC

0.005

[M
[M

2 2 10

h
10
h

gt
gt

en
en

0.002

tl
4
tl

en
en

5 5

m
gm

g
6 6

se
se

jitter jitter 0
0
8 0 8 0

Figure 4: Histogram of the Jitter of Su essfully Figure 5: Histogram of the Jitter of Su essfully
Transmitted Network Layer Segments as a Fun - Transmitted Network Layer Segments as a Fun -
tion of of the Segment Length with Sequential tion of of the Segment Length with SMPT (Con-
Transmission Method (Conden e Level 99% Rel- den e Level 99% Relative Error 1%)
ative Error 1%)

Transmission Segment Loss Mean Jitter these features (the sequential analysis of simula-
Method Probability [MAC pa kets℄ tion output data and parallelization of simulation
Sequential 18.24% 2.438 programs) have been in orporated in AKAROA-2.
SMPT 11.41% 0.761 The work on in reasing fun tionality of this pa k-
age is ontinued both at the Te hni al University
Table 3: Losses and Mean Jitter of Segments with of Berlin, Germany, and the University of Canter-
Dierent Transmission Methods for Ten Wireless bury in Christ hur h, New Zealand.
Terminals (Conden e Level 99% Relative Error
1%)
Referen es
99% onden e level). They an be ompared with
those depi ted in Figure 6 and 7, obtained with the [1℄ Bertsekas, D. and Gallager, R.,Data Net-
relative error of 25% (at the 99% onden e level). works,Prenti e Hall,New Jersey,se ond edi-
One an see the importan e of using simulation tion,1992
results with appropriately small statisti al errors.
As Figure 6 and 7 show, the results with too large [2℄ Proakis, J.G.,Digital Communi ations,
error may be misleading or, at least, in on lusive. M Graw-Hill International Edition,Third
In this ontext, one ould wonder about the red- Edition,USA
ibility of results from sto hasti simulation when
their statisti al error is not assessed at all. [3℄ Fitzek, F., Rathke, B., S hlager, M.,
A.Wolisz,Quality of Servi e Support for
Real-Time Multimedia Aplli ations over
4 FINAL REMARKS Wireless Links using the Simultaneous
MAC-Pa ket Transmission (SMPT) in a
Our results show both importan e of sequential CDMA Environment,Pro . MoMuC 1998,pp
analysis of simulation output data, as the only ef- 367-378,O tober,1998
fe tive way of ontrolling the statisti al error of
the nal results, and usefulness of parallel simu- [4℄ Zorzi, M. and Rao, R.R.,Energy onstrained
lation in MRIP s enario, as a simple and pra - error ontrol for wireless hannels,IEEE Per-
ti al way for speeding up of otherwise very long sonal Communi ations,vol. 4,pp. 27-33,De .
simulations. Automati implementation of both 1997
 0.045
0.045

0.04
0.04

0.035 0.035

0.03 0.03

0.05 0.05
0.025 0.025

Probability
0.04
Probability

0.04

0.03 0.02 0.03

25 0.02
25
0.02 0.02

0.01 20 0.015 0.01 20 0.015

s]
s]

et
et
0

ck
0
ck

15

pa
15
pa

1 0.01 1 0.01

AC
AC

[M
2
[M

2
10

h
10
h

gt
3
gt

en
en

0.005 0.005

tl
4
tl

en
en

5 5

m
5
m

g
g

se
se

jitter 6 jitter 6 0
0
0 7 0
7

Figure 6: Histogram of the Jitter of Su essfully Figure 7: Histogram of the Jitter of Su essfully
Transmitted Network Layer Segments as a Fun - Transmitted Network Layer Segments as a Fun -
tion of of the Segment Length with Sequential tion of of the Segment Length with SMPT (Con-
Transmission Method (Conden e Level 99% Rel- den e Level 99% Relative Error 25%)
ative Error 25%)

[5℄ Ormondroyd, R.F.,Performan e of Low-Rate [12℄ Pawlikowski, K., Yau, V. and M Ni kle, D.,
Orthogonal Convolutional Codes in DS- Distributed sto hasti dis rete-event simula-
CDMA Appli ations,IEEE Transa tions on tion in parallel time streams, Pro . of the 1994
Vehi ular Te hnology,Vol 46,no. 2,May,1997 Winter Simulation Conferen e, pp. 723{730,
1994.
[6℄ IEEE802.11,"Wireless LAN Medium A ess
Control (MAC) and Physi al Layer (PHY) [13℄ S hmeiser, B., Bat h size ee ts in the anal-
Spe i ations",IEEE Stds. Dept.,Jan.,1996 ysis of simulation output, Operations Re-
sear h, Vol.30, no.3, 556{598, May-June 1982.
[7℄ ETSI TC-RES,Radio Equipment and Sys-
tems (RES); High Performan e Radio Lo al [14℄ S hruben, L.W., Dete ting initialization bias
Area Network (HIPERLAN); Type 1,Fun - in simulation output, Oper. Res., Vol.30, pp.
tional Spe i ation,Fran e,De .,1996 569{590, 1982.
[8℄ Heildelberger,P., Statisti al analysis of paral- [15℄ Lee, J.S.R., Pawlikowski K and M Ni kle.
lel simulations, Pro . 1986 Winter Simulation Sequential steady-state simulation: Rules of
Conferen e, 290{295, 1986. thumb for improving a ura y of the nal
results, Pro . European Simulation Simpo-
[9℄ Meketon, M.S. and S hmeiser, B., Overlap- sium, ESS'99, Erlangen, Germany, pp. 618-
ping bat h means: something for nothing ?, 622, 1999.
Pro . 1984 Winter Simulation Conferen e,
227{230, 1984. [16℄ Regents of the University of Cali-
fornia, Berkeley (USA), PTOLEMY,
[10℄ Mota, E., Wolisz, A. and Pawlikowski. K., http://ptolemy.ee s.berkeley.edu
Sequential bat h means te hniques for mean
value analysis in distributed simulation, Pro . [17℄ Ewing, G., M Ni kle, D. and Pawlikowski K.
13th European Simulation Multi onferen e AKAROA-2: Exploiting network omputing
Warsaw, pp. 129{134, 1999. by distributing sto hasti simulation, Pro .
13th European Simulation Multi onferen e,
[11℄ Pawlikowski, K., Steady-state simulation of Warsaw, Poland, pp. 175-181, 1999.
queueing pro esses: a survey of problems and
solutions, ACM Computing Surveys, vol.22,
pp. 123{170, 1990.