Data Trac
Patrik Carlsson 1
, Markus Fiedler 2
and Arne A. Nilsson 3
1
BlekingeInstituteofTechnology,37179Karlskrona,Sweden. E-mail: Patrik.Carlsson@bth.se
2
BlekingeInstituteofTechnology,37179Karlskrona,Sweden. E-mail: Markus.Fiedler@bth.se
3
BlekingeInstituteofTechnology,37179Karlskrona,Sweden. E-mail: Arne.Nilsson@bth.se
Abstract
Recentanalyses of realdata/internet trac indicate
that data trac exhibits long-range dependence as
wellasself-similarormulti-fractalproperties. Byus-
ingmathematicalmodelsofInternettracthatshare
thesepropertieswecanperformanalyticalstudiesof
networktrac. Thisgivesusanopportunitytoanal-
ysepotentialbottlenecks and estimate delays in the
networks.
Processeswithmulti-fractalpropertiescanbemod-
eledbymultiplyingtheoutputofMarkovModulated
RateProcesses(MMRP) [1]eachdened byfour pa-
rameters. The MMRP are easily used in stochastic
uid ow modeling. This model is also suited for
analysis of other trac types e.g. VoIP and thus,
itallowsforintegration of dierent trac types,i.e.
time-sensitivevoicetrac with best-eort data traf-
c. Using this model we can calculate performance
parametersforeachindividualstreamthatentersthe
system/model.
In this paper we show how to construct amulti-
fractalprocessthatismatchedtomeasureddatafrom
MMRPsub processes.
Keywords: Multi-fractal trac, uid ow model,
numericalanalysis,performanceevaluation
1 Introduction
Thereasonforthisresearchisthatwewanttodevelop
a mathematical model for Internet trac, that can
be usedin auid owanalysis. Inearlierpapers[1]
and [2] we demonstrated that the uid ow model
is a usable tool for analytical performance studies.
We have used data processes that had multi-fractal
properties (MFP), and we also expressed desires to
dosimilarstudies basedonprocessesthathavebeen
modeled based on measurednetwork trac. In this
paper we show how a certain class of multi-fractal
processes can be modeled such that they match up
well against real network trac. The tool we now
have permits us to do better capacity management
insteadoftoday'srule-of-thumb.
Sincewearetomatchamathematicalmodelagainst
measureddata,wewillstartbydescribingthemathe-
maticalmodelanditsbasicproperties. Followingthat
wewillverifythismodelagainstsimulationsthathave
beenperformed. The actualmatching oftheparam-
eters basedon real measuredtrac will then be de-
scribed,andnally someconclusionswill bedrawn.
2 Processes description
The mathematical model that is used to generate a
processwithmulti-fractalpropertiesisformedbymul-
tiplying the output of independent 2-state Markov
modulatedrateprocesses(calledsub-processes). The
processwillbeformedas
R (t)= n 1
Y
i=0 R
i
(t) (1)
Here R
i
(t) is the output rate from sub-process i at
time t (data-unit/time-unit). The 2-state Markov-
modulated rateprocessisdepictedin Figure1. This
isawellknownandoftenusedmodel. Themodelwas
originallysuggestedin[4]
2.1 Description of a Sub-Process
Eachsub processisasimplelow-highprocess,asde-
scribedinFigure1. Thesubprocessestransitionma-
trixandratematrixareformedas
M
i
=
i
i
i
i
R
i
=
l
i 0
0 h
i
(2)
ThersttwomomentsoftheMFP rateprocessis
goingtobeusedforthematchingofprocessestoMFP.
Thusitisinterestingtohaveanalyticalexpressionsfor
thetworstmoments.
E[R
i (t)]=
l
i
i +h
i
i
+
(3)
low
high -
i
i
ow 6
rate
h
i
l
i
Figure 1: 2-state Markov Modulated Rate Process,
thei:thasub-process
E
R 2
i (t)
= 2
i
i (l
i h
i )
2
t 2
(
i +
i )
3
t 1
i +
i +
e (
i +
i )t
i +
i
+ (l
i
i +h
i
i )
2
(
i +
i )
2
(4)
A more detailed derivation can be found in [3]. It
is obvious that the equation for the rst moment is
notaectedbyanychangesint. Thesecond-moment
however
lim
t!0 E
R 2
i (t)
=
i
i (l
i h
i )
2
(i+i) 2
+E
R 1
i (t)
2
(5)
lim
t!1 E
R 2
i (t)
= (l
i
i +h
i
i )
2
(i+i) 2
=E
R 1
i (t)
2
(6)
E[Ri2(t)]
E[Ri 1(t)]
10−3t t 103t ∞
0
Time scales Second Moment
Figure 2: Typical shape of the second moment of a
subprocess.
i
=
i
andE[R
i (t)]=1
2.2 Process
Asmentionedbefore theprocessis formed bymulti-
plying theoutputfrom thesub-processes. Thetran-
sitionmatrixisformedusingKroneckeraddition:
M =M M M (7)
R
D
=R
0 R
1
R
n 1
: (8)
Where
AB= AI
B +I
A
B (9)
AB= AI
B
I
A
B (10)
and I
x
is adiagonal matrixwith ones on it and the
samesize asmatrixx,where isdened as
A= 2
4 a
1;1
::: a
1;m
::: ::: :::
a
n;1
::: a
n;m 3
5
nm
B= 2
4 b
1;1
::: b
1;q
::: ::: :::
b
n;1
::: b
p;q 3
5
pq
AB= 2
6
4 a
1;1
B ::: a
1;m B
.
.
. .
.
. .
.
.
a
n;1
B ::: a
n;m B
3
7
5
npmq
(11)
Themoments for theprocess are easily obtained
simplybymultiplicationofthesub-processmoments
E[R (t)]= Q
n 1
i=0 E
R 1
i (t)
(12)
E
R 2
(t)
= Q
n 1
i=0 E
R 2
i (t)
(13)
3 Simulations
To compare the analytical formulas of the rst and
secondmoments,somesimulationshavebeenrun. In
Figure3asetofsimulationsareshownwherewecom-
pareresultsfromsimulationwithresultfromthean-
alytical method. In Table 1 the process parameters
are presented. In Figure 3 the rst and the second
momentareshown. Eachrowcorrespondstoagiven
set ofparameters,startingwith1. Thecolumns cor-
respondstothemoments,columnoneistherstmo-
mentandcolumntwothesecondmoment.
Looking at the rst row thematch is perfect for
bothmoments. Howeverthis process correspondsto
a non varying process, with l
i
= h
i
= 2 hence no
variations. Looking at row two and three, a slight
dierence isdetected. It canprobablybeaccredited
to atoosmall sample space forthe simulation. But
thematchisfairlygoodatleastathighertimescales
forthesecondmoment.
4 Matching
Our developed matching method assumes that the
rstmomenthasbeennormalizedinsuchawaythat
100 105 0
1 2 3
First Moment
100 105
1 2 3 4 5
Second Moment
100 105
0 1 2 3
100 105
1 2 3 4 5
100 105
0 1 2 3
100 105
1 2 3 4 5
Simulated Analytical
Figure3: Acomparisonbetweenmomentsthatwhere
calculated from simulations and moments based on
theanalyticalformulasinEquation3and4
ID l
i h
i
Mean
1 100 0.02 0.02 2 2 2
2 100 0.02 0.02 1 2 1.5
3 83.333 0.03 0.02 1 2 1.4
Table1: Parameters usedinsimulation
themeanoftherstmomentsatdierenttimescales
willbe1.
Whenmatching real-worlddatathere areusually
sometrade-osto bemade. Wehavechosentoover-
estimate the moments. The main reason for this is
thatitisbettertoworkagainstaworstcasescenario.
Anotherconsiderationisthenumberofprocessesthat
areusedtomatchtheprocess,sinceinananalysisstep
tofollow[1]thereisaupper-boundontheprocessing
power. That is if weuse an innite numberof pro-
cesses we would probably get a good match, but it
wouldnotbeusableinfuture work.
When matching weare trying to match therst
twomoments. There are anumber of methods that
couldbeusedtodothis. Inthispaperwehavechosen
to use asimple and heuristic method. The method
worksroughlylikethis:
1. Initialize
~
D 2
Thesecondmoment
~
D 1
Therstmoment
~
T Thetimescales
whereD x
i
denotesthei:thelementinthevector
E[Ri 2(t)]
E[Ri1(t)]
E[Ri1(t)]
10−1t t 101t
Desired i:th process
0 ∞
Time scales Second Moment
Figure 4: How sub-process i is identied from the
desiredsignal
2. For each time scale in
~
T, try to nd a sub-
processthat haspropertiesthatmatchit. Asif
theprocesswereconstant attimescalessmaller
thant. Cf. Figure4.
3. Loop through
~
T starting at the largest time
scale and going down to the smallest, where i
denotesthepositionofthecurrenttimescalein
~
T.
(a) Sett=T
i
, isestimated tobe10t
M 2
=D 2
i M
1
=D 1
i
(14)
These will beused asthedesiredvalue of
the sub-process as t ! 0, since the pro-
cesshasbeennormalizedinsuchamanner
that themean will beoneand remember-
ingthatthelimitvalueofasub-processin
equation5. Thisvaluecanbetoolargefor
onesinglesub-process. Thesolutionisthe
usage ofsub-sub-processes,processesthat
are identical and together achievethe de-
siredpropertiesonagiventimescale.
(b) Estimatehowmanysub-sub-processesthat
areneededtoachievethislimit.
k= l
logM 2
log2 m
(15)
(c) If k <0 thenthis time scale isO.K.goto
3.
(d) Determinethelimitforasub-sub-process
O 2
= k p
M 2
O 1
= k p
M 1
(16)
(e) Basedonthesymmetryfrom==2=
an expressionfor landhcanbeobtained
inthefollowingmanner:
L= 2 p
O 2
(O 1
) 2
(17)
l= M
1
i
L
2
(18)
h= M
1
i +
L
2
(19)
(f) The parameters that identify sub-process
i are known (the output rates in the two
states, the transition rates and the num-
berofsub-sub-processesthatformthesub-
process). Create twovectors
~
M 2
i and
~
M 1
i
thatspanalltimescalesdenedby
~
T from
these parameters. These vectors specify
the moments for sub-process i. Store the
parametersneededtocreatethesevectors.
(g) Remove the inuence on the desired pro-
cess that this sub-process has by simply
dividing it out, element-by-element. j is
thesizeofthe
~
T.
~
D 1
= h
D 1
1
M 1
1
; :::; D
1
j
M 1
j i
(20)
~
D 2
= h
D 2
1
M 2
1
; :::; D
2
j
M 2
j i
(21)
4. Now the matched process can be created by
combining the sub-processesusing element-by-
elementmultiplication using Equation 3and 4
from the stored parameters. Recreate themo-
mentvectorsforeachsub-processcallthem
~
P x
i
where i indicate that this is sub-process i and
x indicate thex:thmoment (one ortwo). The
momentswillthenbe
~
M 1
= 2
6
4 P
1
1;1
P 1
2;1 :::P
1
j;1
;
:::;
P 1
1;n 1
P 1
2;n 1 :::P
1
j;n 1 3
7
5 (22)
~
M 2
= 2
6
4 P
2
1;1
P 2
2;1 :::P
2
j;1
;
:::;
P 2
1;n 1
P 2
2;n 1 :::P
2
j;n 1 3
7
5 (23)
5. Repeat step3as many timesasdesired (could
bebasedonmaximalerrorormaximalnumber
ofprocesses).
Now itis possibleto listtheparametersthat are
neededto matchthedesiredinput. Whencomparing
amatchedprocesswithameasuredprocesswedene
theerroras
!
Error=
"
D
2
1 M
2
1
D 2
1
; ::: ;
D
2
j M
2
j
D 2
j
#
(24)
4.1 Example 1: Simulated data
InFigure5anexampleofthematchingisshown. Here
i i
1 512 512 0.5 1.5
2 128 128 0.5 1.5
3 32 32 0.5 1.5
4 8 8 0.5 1.5
5 2 2 0.5 1.5
Table2: Parametersusedtocreatethedesiredsignal
thatis matchedinExample1.
byvesub-processeslistedin Table2. Therstmo-
ment is aperfect match, the second moment is over
estimated. Looking at theerror the overestimation
isnotthatlarge.
10−5 10−4 10−3 10−2 10−1 100 101 102 103
0.9 0.95 1 1.05 1.1
First Moment
Measured Matched
10−4 10−3 10−2 10−1 100 101 102 103
1.5 2 2.5 3
Second Moment
Measured Matched
10−4 10−3 10−2 10−1 100 101 102 103
0.02 0.04 0.06 0.08 0.1 0.12 0.14
Second Moment Error
Figure 5: Comparison between a simulated process
andacreatedprocess. Example1.
4.2 Example2: Uplinkon small work-
group switch
Thedataused herecame fromamid-size workgroup
switch(Cisco2948G).Therewereabout36hostscon-
nected to it, distributed into twocategories: depart-
ment computersand 16 computersaccessibleto stu-
dents. Themeasurementswereperformedduringthe
summer of 2001, from the 7:th to the 30:th of July,
whichresultedinmorethan210 6
samples. Themea-
surementfocusedonthe interfacethatcorresponded
to the uplink of the switch. Thedata wascollected
usingasimpleSNMP (SimpleNetwork Management
Protocol) pollingtool(developed during 2001-2002).
Thetoolpolledthedeviceroughlyonceeverysecond,
withtheodddelay(probablyduetoprocessingtimes
in the device and intermediate devices). This jitter
isofnogreaterimportancesincethevaluesarerecal-
culated andstoredasrates,thusifthemeasure took
100 101 102 103 104 105 200
400 600 800 1000 1200
Second Moment
Measured Matched
100 101 102 103 104 105
0.2 0.4 0.6 0.8 1
Second Moment Error
Figure 6: Comparison between a measured process
andacreatedprocess. Example2.
LookingatFigure6themeasuredprocessrstmo-
ment behaves as expected, but the second moment
seemstolackofaat,orsemi-atsection. Onepossi-
blereasonforthisbehaviorcouldthatthetracload
was low (measured during the summer), and most
of the trac that were present came from comput-
ers talking to computers in a (almost)deterministic
manner, e.g. WINS, NFS, etc. This type of source
generatestrac at discreteintervals, andthepacket
sizesarealmostalwaysthesame.
Turning the focus towardthe matched process a
under-estimationis noticed in the rst point. A so-
lution to this problem, and perhaps a renement of
themethod,wastoreapplythemethodtothewhole
set of data that remained after the rst cycle. By
doing this aeven better match wasachieved in this
point, butstill under-estimated(the otherpointsre-
mained unchanged sincethe method found that the
pointsalreadywereover-estimated,k<0). Oneway
toover-estimateall-pointswouldbeto addaprocess
with a time scale larger than the largest time scale
andusing ittoshiftthe entireprocessupward. This
wouldhoweverincreasetheerroratalltimescales.
Toseeifitwaspossibletogetevenbettermatches
themethod wasmodiedby setting k =1in Equa-
tion 15 and run it several times. The result was a
dramaticallyimprovedmatch,thecostwasansignif-
icantincreaseofprocesses.
5 Conclusions
Inthispaperwehavepresentedacrudemethodthat
allows us to match process parameters against real
network trac. It is possibleto get abetter match
usingmoreprocesses,this,however,willcostinlater
100 101 102 103 104 105
200 400 600 800 1000 1200
Second Moment
Measured Matched
100 101 102 103 104 105
0.05 0.1 0.15 0.2 0.25 0.3 0.35
Second Moment Error
Figure 7: The resulting match after running the
method5times,modifyingthemethod
possibleto get afairly good match using onlya few
processes. We need ameasurementtool that allows
ustomeasureatsmallertimescalesthanonesecond.
SNMPmeasurementseemsnottobeabletosupport
this, in fact severalmeasurementshave showedthat
theMIB (ManagementInformation Base)is notup-
dated in a correct way, but rather in 10 seconds in-
tervals (thisbehaviorwasmostnotably noticed in a
well-knownoperatingsystem).
References
[1] P. Carlsson and M. Fiedler. Multifractal prod-
uctsofstochasticprocesses: Fluidowanalysis.
Proceedings of 15th Nordic Teletrac Seminar
(NTS-15),Lund,Aug.22-24,2000.
[2] P.Carlsson,M.FiedlerandA.Nilsson.Voiceand
multifractaldata intheInternet.Proceedings of
26th IEEEConferenceon Local ComputerNet-
works(LCN2001),Tampa,Nov.15-17,2001.
[3] P. Carlsson, M. Fiedler and A. Nilsson. Mo-
ment analysis of sub-processes of multi-fractal
sub-processes,Internal Report,2002.
[4] P. Mannersalo, I. Norros, and R. Riedi. Multi-
fractal products of stochastic processes: A pre-
view.
COST-257 Technical Document 257TD(99)31.
http://nero.informatik.uni-wuerzburg.de/
cost/Final/TDs/257td9931.pdf