A new approach to model the stationary behavior of TCP connections

Abstract

ttb.urflcr-In ibis pqw, we outllnc a nictliodology that can lie applied lo rirodel the hehilvior of TCP h w s . The proposed trethodology stcins frunr B Mnrltuvinii model of n single TUP source, nnd cvcntriully coiisidcis the su. pci~osition aiid iiitcracllon uf scvcrnl such suiircccs rising stniitlanl qiicucirrg aiialysk twhniqtlu. Oiir approncli nllows thc evaluntion o f such perforiiiincc indica m tliraughpiit, qiieiiclng 11cli1y mil pnckct lws of TCP flows. 'I'he restilts obtained through ~U F mndcl are vnlidnted by iiieaiis of s1111uIa-liuii, under several topology nnd trnftic settings. I . INTKODUC'l'IONAccording to recent estitiintcs 95% of the h-aliic c;trriecl toclay over widc-arm 11' iictworks uses TCP as transpo1.t protocol, which ainouiils to 80% of tlic ovcrall end-lo-cnd flow count. These figures alonc highlight the key rnle that 'I'CP plays in delivering a rcliable servicc tu the most comnioii iicrwork applications such as ernail programs and Web browsers. Diiring the innny years of "Iionorablc" scrvice, TCP tias undergono scvcral alI.clations that Iiave in& it mol-c robust lo cliangitrg traffic patteri~s and to nctwtirk cimgcstion. However, the rcsulliiig co~riplexity OT the algoIittims governing TCP's flow control dynaiiiics has so fiir dcficd any nccuralc analylic mprcsontation.In recent years, several clrwts havc bceii dcvoted to modcliitg such a coinplcx protocol as TCP. Some papers Iiwc tricd to capture the essential TCP dynamics through closed-form exprcssions. Lakshman and Madhow [ l ] and Kumar [Z] use Markovian analysis to dcvclnp il closcil-form expression for the tl~rc~ughput or TCP connections by observing thc cyclical cvolution ofthc TCP transmissiori window; lhc latter work introduces sonic c.xtcnsioiis h r scvcral vcrsioris of X P , incorpor;iting such features EIS coarse timers, fast rctransmit and has1 recovcry. A fluid rnotlcl o f i i single TCP coniicction sharing a link with bmkgruuiid, noii-'I'CP traffic was proposed by hltman ct al. 131 in R tiyhrid approach combining both analytical and cxpcrimenta1 work; a q was done in other, inme recent studies, their modcl proceeds tbmugb a stochastic aiirilysis of the dynamic behavior of 1'ClP's congcstion window sixe: specifically, they aim at PTOviding aa cxprcssion for the throughput nf a single crmnechn. Similarly, Mathis et til. [4j fncusctl on ttie stochastic bchavior of tlic Congestion Avoidancc mcchatiism, deriving an cxpression for thc. ttiroiighput that was tlrcn npplietl to study tlic behavior scvcml flavors of ' K P so~irccs and qucucing tcctiniqucs. Rcccritly, Padhye et al. I S ] havc devdopccl a steady-state niodd rhnc appmximatcs the thronghpiit of bulk TCF jlows as R fiinction of loss rate nnrl round trip lime, coinparring their estiinales witli rc.nl-lifc traccs of ' K P tml'lic. Notably, smnc rescnrclicrs [GI, 1:7] have chosen a different ap-.I ills work \VIIS ruplJorietl by ithe ~mlinti Miiiisrry for Universiiy xitl SciciitificI<cse:u-ch, proach wbcrehy tbc bbserviition of "actual" TCP races was the foundation of cmp...

Efficient steady-state analysis of second-order fluid stochastic Petri nets

“…Works in [1,5] cope with finite TCP connections; their peculiarity is the analysis of connections that exhibit an on-off behavior, following a Markov model. In these works, the description of the protocol behavior is decoupled from the description of the network behavior and the interaction between the two sub-models is handled by iterating the solution of the sub-models until the complete model solution converges according a fixed point algorithm.…”

“…In these works, the description of the protocol behavior is decoupled from the description of the network behavior and the interaction between the two sub-models is handled by iterating the solution of the sub-models until the complete model solution converges according a fixed point algorithm. In [1] the GSPN formalism was used for the specification of the TCP and the network behavior, while in [5] simple analytical models were developed.…”

On the Use of Queueing Network Models to Predict the Performance of TCP Connections

“…Markovian modeling is used in [7], that presents Markov reward models of several TCP versions on lossy links. Also the work in [8,9] is based on a Markovian approach. The novelty in these works is the consideration of connections that switch on and off, following a two-state Markov model; however, the main performance figure is still the steady-state TCP connection throughput, since the model cannot provide estimates of completion times.…”

A detailed and accurate closed queueing network model of many interacting TCP flows

“…The approach followed in [10], [11] is based on Markovian modeling. The work in [10] assumes an underlying ATM network, and is limited by the solution complexity when the number of TCP connections grows.…”

“…The work in [10] assumes an underlying ATM network, and is limited by the solution complexity when the number of TCP connections grows. An innovation introduced in [11], that does not assume an underlying ATM network, is the presence of non-greedy connections.…”