行政院國家科學委員會專題研究計畫 成果報告
子計畫四:All-IP 網路 End-to-End 品質管理之研究(II)
計畫類別: 整合型計畫
計畫編號: NSC92-2219-E-004-004-
執行期間: 92 年 08 月 01 日至 93 年 07 月 31 日
執行單位: 國立政治大學應用數學學系
計畫主持人: 陸行
報告類型: 完整報告
處理方式: 本計畫可公開查詢
中 華 民 國 93 年 8 月 18 日
!"#$% &'()*'+)*,-+ ./(NSC 92-2219-E-004-004 0 12(9238415693374315 789(: ;<789( =>?@ABCDEFG( )HIJKLÆMNOPQ )HR:STJKLÆMNOPQ )JUVW XNOYZ[E\]^PQ )V-_ÆIÆ`PQ 0 a(bRBcd efg9338425
¡¢£¤¥¦§¨©ª«¬¡Æ¯
MathematicalModels of Pareto Optimal Path Selection on All-IP Networks
: NSC 92-2219-E-004-004 : 920801930731 : Æ
1 Abstract
We present an approach for the fair resource allo-cation problem and QoSroutingin All-IP networks thatoermultipleservicestousers. Theobjectiveof theoptimizationproblemistodeterminethe amoun-tofrequiredbandwidth foreach linkand eachclass to maximize the sum of the users' utility. In this work, wefocuson approachesthat, while allocating bandwidth,attempt to provide aproportionallyfair treatmentofallthecompetingclasses. First,wewill showthatanachievementfunctioncanmapdierent criteria subject to various utility onto anormalized scale. It may be interpreted as a measure of QoS (Quality of Service) on All-IP networks. Using the bandwidth allocation model, we can nd a Pareto optimalallocationof bandwidthonthenetwork un-der a limited available budget. This allocation can provide the so-called proportional fairness to every class, that is, thisallocationcanprovide thesimilar satisfactiontoeachuser. Next,wepresentarouting schemeunderconsiderationofthedelay. Suchan op-timal path provides the end-to-end QoS guarantees to each user. Finally, a numericalexample is given toillustrate howtosolvethefairresourceallocation problemandhowto modifythenonlinearparts. (Keywords: multiple-objectiveproblems,routing, achievementfunction,proportionalfairness,delay, Paretooptimal,orderedweightedaveragingmethod, fairbandwidthallocation)
e]hi Æ ÆÆ !All-IPÆ"#$%&Æ'( )*+,-./Æ01 2&3456 789) :)*;<=>?@ABCD=> E+FGHIJKLMNÆEO&PQ All-IPÆ&RS)*;<TUFVWXYZ [\]^_`"aKLbc+de fgh(i - jk(achievement function)lm0!no N pqrst\uvwTUxt!nFV+y z {|A}~lAll-IPÆ (end-to-end)a=>+ \ fg}DZ Q xt+wQ?pq%&Æp (j :¡xtp( .t+fg"¢£¤¥¦ ;<§0 "¨©p(EªB «j+ (jkl: !"m nopqnrsn tu#vwxyzn{x#|}~nn {xn ) 2 Introduction
Packet switchednetworks suer threemajor quality problemsin oeringtime-sensitiveservices: long de-laytime,jitter,packetloss. TheUniversalMobile T-elecommunicationsSystem(UMTS) [1]hasspecied
[1: UMTSServiceClasses
Traffic Classes
Examples
of Applications
Sensitivity
to Jitter
Sensitivity
to Delay
Sensitivity
to Packet Loss
Conversational
VoIP
high
high
low
Streaming
VoD
high
high
low
Interactive
WWW, Telnet
low
low
high
Background
E-mail, FTP
very low
low
high
fourdierenttraÆcclassesaccordingtotheirquality of service (QoS) requirements for dierent applica-tionsasTable1shows. Dierentpeoplehavedierent expectationstothenetworkQoS.Thereareanumber of characteristics that qualify QoS, including mini-mizing delivery delay, minimizing delay variations, providingconsistentdatathroughputcapacity.
QoSrouting concernsthe selection of apath sat-isfying the QoS requirements of a ow. The path selected most likely is not the traditional shortest path. Dependingonthespecicsandthenumberof QoSmetricsinvolved,computationrequiredforpath selection can become prohibitively expensive asthe network size grows. The path selection process in-volvestheknowledgeofthe ow'sQoSrequirements and characteristics and (frequently changing) infor-mationon the availability of network resources (ex-pressedin termsof standard metrics, e.g., available bandwidthanddelay). Resourceallocationdecisions areconcernedwiththeallocationoflimitedresources soastoachievethebest systemperformances.
In a multi-objective decision-making situation in theabsenceof uncertainty weoftensearchfor Pare-to optimal solutions. One scheme for dealing with multi-objective models that permits more balanced handlingoftheobjectivesissimplytocombinethem in aweightedsum. Multiple objectivefunctions can becombinedintoasinglecompositeoneto be max-imized bysumming objectiveswith positiveweights onmaximizing and negativeweightson minimizing. Ifthecompositeistobeminimized,weightson max-imizing objectivesshould be negative, and those on minimizingshouldbepositive. Signsorientall objec-tivesin thesamedirection,andweightsre ecttheir relativeimportance. Ifasingleweighted-sum objec-tive model derivedfrom amulti-objective optimiza-tionproducesanoptimalsolution,thesolutionisan
Inthiswork,weusethemethodofweightedsumsto solveourproblems.
We dealwith the problem of dimensioning band-widthforelasticdataapplicationsinpacket-switched communication networks, which can be considered as a multiple-objective optimization model. In our work,wewillfocusonthefollowingsubjects: (i)How do we transform the dierent criteria measurement onto anormalizedscale? (ii)Howdo weallocate re-sourceswithproportionalfairnessandndarouting schemeonAll-IPcommunicationnetworks? (iii)How dowemodifythenonlinearmultiple-objective prob-lemsassolvableMixed-Integerprogrammingmodels?
3 Achievement Function
Inordertotransformthedierentmeasurements on-to anormalizedscale, weconstructtheachievement function
i
for each criteria i which can be viewed as an extension of the fuzzy membership function in terms of a strictly monotonic and concave utili-ty function as shown in Figure 1. We assume that the decision maker species requirements in aspira-tionandreservationlevelsbyintroducingdesiredand requiredvaluesfor several outcomes. Depending on thespeciedaspirationandreservationlevels,a
i and r
i
, respectively, weconstruct ourachievement func-tionofz i asfollows: i (z i )=log z i r i ; where= a i r i : (1) Formally,wedene i
()overtherange [0;1),with
i
(0)= 1and 0 i
(0)=1. Dependingonthe spec-ied reference levels, this achievementfunction can be interpreted asa measure of thedecision maker's satisfactionwith the value of thei-th criteria. It is astrictly increasingfunction of z
i ,havingvalue1if z i = a i , and value 0 if z i = r i . The achievement function canmap thedierentcriteriavaluesontoa normalizedscaleofthedecisionmaker'ssatisfaction. Moreover,thelogarithmic achievementfunction will beintimately associatedwiththeconceptof propor-tionalfairness(see[6]and[8]). Wewillformulatethe mathematicalmodelofthefairbandwidthallocation byusingtheachievementfunction.
satisfaction
i
a
i
r
z
i
1: TheGraphofanAchievementFunction i
(z i
)
4 Formulation of the Band-width Allocation Model with Proportional Fairness
Given a network topology G =< V;E >, where V andE denotethesetofnodesandthesetoflinksin the network respectively. There is given a set S of m classes, i.e., jSj = m. We denote by S
i
aset of sessions in class i. There is also giventhe maximal possiblenumberK
i
ineachclassi,thatisjS i
j=K i
. Wewillgetthefollowingmathematicalmodel(MP1):
Maximize m X i=1 w i i Subjectto X e2E e x e =B X i X j i j (e) i j =x e ; 8e2E X i (K i c i + i )=B i j b i ; 8j2S i ; fori=1;:::;m x e U e ; 8e2E i =it i m X k =1 d k i ; 8i=1;:::;m t i d k i f k (x); 8i; k=1;:::;m d k i 0; 8i; k=1;:::;m i j X e e i j (e)=c i ; 8j2S i ; fori=1;:::;m i 1 = i 2 == i K i ; 8i=1;:::;m x e 0; 8e2E i j 0; 8j2S i ; fori=1;:::;m i j (e)=0or1; 8e2E; t i unrestricted,8i=1;:::;m; wherew m = m ,w i = i i+1 fori=1;:::;m 1, i
2(0;1)is givenforeachi, and P m i=1 i =1. The individualfunction i
istherstisumoftheordered multipleobjectivefunctions
i
intheallocation pat-ternx=fx
e
je2Egandthebandwidth i
allocated toclassi. Here,weletK
i
in(MP1)beaxednumber forthediscussionunder deterministicassumptionof feasibilityof(MP1). Ingeneral,K
i
mayberandom which isbeyond scopeofthethesis.
5 Modications of Nonlinear Parts
Werewrite(MP1)asthefollowingmodel(MP2).
Maximize m X i=1 iw i t i m X i=1 m X k =1 w i d k i subjectto X e2E e x e =B X i X j A i j (e)=x e ; 8e2E X i (K i c i + i )=B x e U e ; 8e2E d k i 0; 8i; k=1;:::;m A i j (e)+b i M i j (e); 8e2E; 8j2S i ; fori=1;:::;m A i j (e)M(1 i j (e)); 8e2E; 8j2S i ; fori=1;:::;m
e e A i j (e)=c i ; 8j2S i ; fori=1;:::;m t i d k i z i 1 ^ f i (0)+z i 2 ^ f i (b i;1 )+z i 3 ^ f i (1) +z i 4 ^ f i (b i;2 )+z i 5 ^ f i (b i;3 ) z i 6 ^ f i (10) z i 7 ^ f i (b i;4 ) z i 8 ^ f i (M i ); 8i; k=1;:::;m i =z i 2 b i;1 +z i 3 +z i 4 b i;2 +z i 5 b i;3 +10z i 6 +z i 7 b i;4 +z i 8 M i ; fori=1;:::;m z i 1 y i 1 ; fori=1;:::;m z i k y i k 1 +y i k ; 8k=2;:::;7; i=1;:::;m z i 8 y i 7 ; fori=1;:::;m P 8 k =1 z i k =1; fori=1;:::;m P 7 k =1 y i k =1; fori=1;:::;m y i k =0or1;8k=1;2;:::;7; i=1;:::;m z i k 0;8k=1;2;:::;8; i=1;:::;m x e 0; 8e2E i j (e)=0or1; 8e2E; 8j2S i ; fori=1;:::;m A i j (e)0; 8e2E; 8j2S i ; fori=1;:::;m t i unrestricted,8i=1;:::;m; wherew m = m ,w i = i i+1 fori=1;:::;m 1, i
2(0;1)isgivenforeachi, and P m i=1 i =1. 6 Conclusions
Inthiswork,wepresentanapproachforthefair re-sourceallocationproblemandQoSroutinginAll-IP networksthat oermultiple servicestousers.Users' utilityfunctionsaresummarizedbymeansof achieve-mentfunctions. First,wendthat theachievement functioncanmapdierentcriteriaontoanormalized scale. Theachievementfunctionalsocanworkinthe Ordered WeightedAveragingmethod. Moreover,it may be interpreted as a measure of QoS on All-IP networks.Usingthebandwidthallocationmodel,we canndaParetooptimalallocationx
ofbandwidth onthenetworkunderalimitedavailablebudget,and this allocationcanprovidetheso-called proportion-al fairness to everyclass i. That is, this allocation canprovidethesimilarsatisfactiontoeachuserinall classes. Wealsondthebandwidthallocatedtoeach classi. Moreover,weobtainthemaximalrate,which the link can oer to each class. Next, we present arouting scheme under considering thedelay. This schemeaimsatseekingapathforwhichtheresidual
maximalrate(i.e.,afterestablishingthenew connec-tion) of its bottleneck link is maximal. This opti-mal path provides the End-to-End QoS guarantees toeachuser.
[1] \Enabling UMTS/Third Generation Services andApplications",UMTSForun ReportNo.11, Oct.,2000.
[2] A. W. Berger, and Y. Kogan, \Dimensioning BandwidthforElasticTraÆcinHigh-Speed Da-taNetworks",IEEE/ACMTransactionson Net-working,Vol.8,No.5,Oct.2000.
[3] J.Gozdecki, A. Jajszczyk, and R. Stankiewicz, \QualityofServiceTerminologyinIP Network-s",IEEE Communications Magazine,no.3,pp. 153-159,Mar.2003.
[4] P. V. Hentenryck, ILOG OPL Studio 3.5: The OptimizationLanguage,MassachusettsInstitute ofTechnology,April2001.
[5] Y. N. Lien, H. C. Jang, T. C. Tsai, and H. Luh,\BBQ:AQoSManagementInfrastructure forAll-IPNetworks",Communicationsof Insti-tuteof Information andComputing Machinery: MobileCommunicationsandWirelessNetworks, Vol.7,No.1,pp.89-115,Mar.2004.
[6] W. Ogryczak, T.
Sliwinski and A. Wierzbick-i, \Fair Resource Allocation Schemes and Net-work Dimensioning Problems", Journal of T-elecommunicationsandInformationTechnology, pp.34-42,3/2003.
[7] A.Orda,\RoutingwithEnd-to-EndQoS Guar-antees in Broadband Networks", IEEE/ACM TransactionsonNetworking,Vol.7,No.3,June 1999.
[8] M.Pioro,G.Malicsko,andG.Fodor,\Optimal Link Capacity Dimensioning in Proportionally Fair Networks", NETWORKING 2002, LNCS 2345,pp.277-288,2002.