0 20 40 60 80 100 120 140 160 180 200
0 500 1000 1500 2000 2500 3000 3500
Link Capacity (Mbps)
WWW Delay (msec) WFQ Bandwidth as a function of WWW Delay
Figure10.4: WFQCapacityvsEmailandWWWDelayforCase4: increased
Emailand WWW variance
analysis is not able to capture the eects of this nonlinearity hence the
pre-dicted sensitivity coeÆcients for Email and WWW are not accurate. This
inaccuracy inturnaectsthe varianceresults. However wecanobservefrom
the simulationthat Emailis nowthe most signicantfactor, contributingto
about50%ofthevariance,followedbyWWWtraÆc. ForCBQ,Table 10.14
shows thatemail andWWW are themost signicantfactors contributingto
the varianceincapacity, togetheraccounting foralmostallof thevariancein
capacity.
In terms of the sensitivity however, voice is still the highest followed by
emailthen videoandWWW. ThePQresultsare shown inTable 10.15. We
observethesametrendsasforCBQinthevariancecontributionofeachtraÆc
Sensitivity(Mbps/msec) 79.3 5.13 12 3.75
p
Variance(Mbps) 23.2 9.47 497.94 392.25
% Variance 0.13 0.02 61.5 38.35
Table10.14: CBQSimulationResultsforCase4: increasedEmailandWWW
variance
Parameter Voice Video Email WWW
Sensitivity(Mbps/msec) 32.7 3.04 12 3.85
p
Variance(Mbps) 11.6 6.32 506.85 402.74
% Variance 0.03 | 61 38.97
Table 10.15: PQSimulationResultsforCase 4: increased Emailand WWW
variance
type although the sensitivity indices are dierent. Voice has the highest
sensitivityindex followed by email,thenWWW andVideo. ForFIFO,voice
and video are the most signicant factors both in terms of sensitivity and
variance in capacity as shown in Table 10.16. This is because only a small
Parameter Voice Video Email WWW
Sensitivity(Mbps/msec) 2820 182 0.679 0.184
p
Variance(Mbps) 825.8 336.98 64.17 68.65
% Variance 84.8 14.1 0.5 0.6
Table 10.16: FIFO Simulation Results for Case 4: increased Email and
WWW variance
proportion of samples from the email and WWW distributions (0.5% and
0.25%respectively)aresmallenoughtobetheminimumvaluethatdetermine
the capacity.
The four cases in this section have validatedthe methodology and
analy-sis that can be used to perform uncertainty and sensitivity analysis for the
four traÆc handling schemes. Numerical simulations can be used to
per-formthe sensitivity analysiswhilethe theoreticalanalysis isvalidonlywhen
the capacity-delay function is linear. More work is needed to extend and
generalize the theoretical formulation tocoverall cases.
In this chapterwe have shown how astochastic formulationof the delay
re-quirements can be used to provide some understanding of how the capacity
requiredbydierenttraÆchandlingschemesisaectedbyuncertaintyinthe
values ofthe maximumdelaybounds. Wehavedeveloped anddemonstrated
amethodology forperforminguncertaintyand sensitivity analysiswhich can
be used to capture how uncertainty in delay bounds translates into
uncer-tainty inthenetworkcapacity. Apurely analyticsolutionhas beenobtained
with the accuracy being largely dependent on the linearity of the
capacity-delayfunction. Forcaseswherethefunctionsarenotlinear,numericalMonte
Carlo simulations were used to provide accurate results. In the next
chap-ter we conclude by discussing the relevancy of our results in the context of
network planningand design.
Conclusion
11.1 Implications of Results on Network
Ar-chitectures
Weconcludebydiscussing theresults obtainedinthe contextofcurrent and
proposedapproachestothe useoftraÆchandlingformulti-servicenetworks.
Best-Eort Network
Thebest-eortnetworkusesFIFOinboththeedgeandcoresothatthe
delay guarantees are uniform across all traÆc types. From our results
wend that tosupport theQoS of delay-sensitive applicationssuchas
voicerequires abundantnetwork capacitywhen the voicetraÆc shares
aqueuewithburstytraÆcsuchasemailandWWW.Whenthereisno
bursty traÆcthen anall-FIFOnetworkperformsjustaswell asa
non-FIFO network. This may seem to suggest the use of separate queues
and links (in essence a separate network) for delay sensitive traÆc to
isolateitfromthe burstynon-delay sensitivetraÆc. The appealofthe
best-eort network lies in its simplicity and if network capacity is not
a constraint, then itmay stillbe the network of choice for some.
Class-BasedNetwork
In a class-based network, ows are grouped into distinct classes and
resources suchas queues and linkbandwidth are allocatedto the class
as whole. TraÆc handlingapproaches for this type of network include
Class-Based Queueing, Priority Queueing and Class-Based Weighted
Fair Queueing among others. Our results show that using class-based
can meet the delay guarantees of all traÆc types with minimal
band-width. Class-based handling in the edge with best-eort handling in
the core requires moderatecapacity while using best-eortin the edge
increases the bandwidth signicantly. The complexity of class-based
handling may range from fair to extreme depending on the exact
im-plementation.
Flow-Based Network
In a ow-based network, each ow is allocated its own dedicated
re-sources and as such managing the network may prove to be complex
when there are numerous ows. Using ow based handling such as
WFQ requires minimal capacity and any combination of ow-based
handling and class-based handling does not increase the bandwidth
signicantly. Using ow-based handlingwithbest-eorthandlingin
ei-ther the edge or core of the network requires more abundant capacity.
Figure 11.1summarizesthecapacityrequirementsofcombinationsofedge
and core traÆc handlingmechanisms.