Comparative study of High-speed Linux TCP Variants over High-BDP Networks MohamedA.Alrshah1,a,MohamedOthman2,a,BorhanuddinAlib,ZurinaMohdHanapia aDepartmentofCommunicationTechnologyandNetwork,UniversitiPutraMalaysia,43400UPM,Serdang,SelangorD.E.,Malaysia bDepartmentofComputerandCommunicationSystemsEngineering,UniversitiPutraMalaysia,43400UPM,Serdang,SelangorD.E,Malaysia. Abstract 6 TransmissionControlProtocol(TCP)hasbeenprofuselyusedbymostofinternetapplications. Since1970s,several 1 0 TCPvariantshavebeendevelopedinordertocopewiththefastincreasingofnetworkcapacitiesespeciallyinhigh 2 BandwidthDelayProduct(high-BDP)networks. IntheseTCPvariants,severalapproacheshavebeenused,someof theseapproacheshavetheabilitytoestimateavailablebandwidthsandsomereactbasedonnetworklossand/ordelay n a changes. Thisvarietyoftheusedapproachesarisesmanyconsequentproblemswithdifferentlevelsofdependability J andaccuracy.Indeed,aparticularTCPvariantwhichisproperforwirelessnetworks,maynotfitforhigh-BDPwired 6 networksandviceversa.Therefore,itisnecessarytoconductacomparisonbetweenthehigh-speedTCPvariantsthat 2 haveahighlevelofimportanceespeciallyafterthefastgrowthofnetworksbandwidths.Inthispaper,high-speedTCP variants,thatareimplementedinLinuxandavailableforresearch,havebeenevaluatedusingNS2networksimulator. ] I ThisperformanceevaluationpresentstheadvantagesanddisadvantagesoftheseTCPvariantsintermsofthroughput, N loss-ratioandfairnessoverhigh-BDPnetworks. Theresultsrevealthat,CUBICandYeAHovercometheotherhigh- s. speed TCP variants in different cases of buffer size. However, they still require more improvementto extend their c abilityto fullyutilize the high-speedbandwidths,especiallywhenthe appliedbufferis near−zero orless than the [ BDPofthelink. 1 Keywords: LinuxTCP,High-BDP,CongestionControl,Throughput,LossRatio,FairnessIndex. v 4 3 5 1. Introduction plemented in several operating systems and examined 3 inrealenvironment. Withtheadvancementinnetwork 0 Transmission Control Protocol (TCP) is commonly technology, TCP faced many new scenarios and prob- . usedbymostofInternetapplicationsandbecomesone 0 lems, such as network congestion, under utilization of 1 of the two original components of the Internet proto- bandwidth, unfair share, unnecessary retransmission, 6 colsuite,complementingtheInternetProtocol(IP),thus outoforderdelivery,non-congestionloss. Allofthese 1 the entiresuite is knownas TCP/IP. TCP providessta- problems encouragedresearchers to review the behav- : v bleandreliabledeliveryofdatapacketswithoutrelying iorofTCP.Inordertosolvetheseproblems,manyTCP i onany explicitfeedbackfromthe underlyingnetwork. X variants have been developed. Each TCP variant has However,it relies onlyon the two endsof the connec- been designed to solve certain problems, some try to r tion which are sender and receiver. That is why TCP a surviveoveraveryslowandcongestedconnections,and isknownasend-to-endorhost-to-hostprotocol. Inthe sometrytoachievehigherthroughputtofullyutilizethe last couple of years, TCP is profusely used by major high-speedbandwidths,whilesometrytobemorefair. Internetapplicationssuchasfiletransfer,email,World- Infact,theyaremostlydifferentfromeachothersothat Wide-Webandremoteadministration. categorizesthemintohigh-speed,wireless,satelliteand The first idea of TCP had been presented by lowpriority. Indeed,aparticularTCPvariantwhichis CerfandKhan (1974). Thereafter, TCP has been im- properforwirelessnetworks,maynotfitforhigh-BDP wirednetworksandviceversa. 1Correspondingauthors: E-mailaddresses:[email protected](MohamedAlrshah), Therefore, it is necessary to conduct a comparison [email protected](MohamedOthman). betweenTCPvariantsthataredesignedforhigh-speed 2TheauthorisanassociateresearcherattheComputational Sci- networksto show the advantagesanddisadvantagesof enceandMathematical PhysicsLab, Institute ofMathematical Sci- ence,UniversitiPutraMalaysia. eachTCPvariant.Inthispaper,ScalableTCP,HS-TCP, PreprintsubmittedtoJOURNALOFNETWORKANDCOMPUTERAPPLICATIONS October13,2016 BIC, H-TCP, CUBIC, TCP Africa, TCP Compound, TCP Fusion, NewReno, TCP illinois and YeAH have beenevaluatedusingNS2networksimulator. Thisper- formanceevaluationpresentstheadvantagesanddisad- vantagesof the comparedTCP variants and shows the differencesbetweenthemin termsof throughput,loss- ratio and fairness over high-BDP networks. As well as, it presents and explains the behaviors of the com- paredTCPvariants, showsthe impactsof the used ap- proaches, and arranges the thoughts. Thus, this paper mayhelptheresearcherstoimprovetheperformanceof theexistingTCPvariantsbycuttingdowntheeffortof comparingtheexistingprotocolsinordertoimproveit tofitthenewgenerationofthenetworks. Therestofthispaperisorganizedasfollows:Section 2presentsthemotivationsbehindthiswork,challenges andpreviousworks. While,Section3presentstheper- formanceevaluationofhigh-speedTCPvariantsandex- plainstheexperiments’setup,networktopology,perfor- mancemetrics, resultsanddiscussion. Finally,Section 4concludesthepaperwithsomefinalcomments. 2. Motivations,ChallengesandPreviousWorks Figure1:TheclassificationandevolutionofvariantsofTCPconges- Therapidgrowthofnetworktechnologiesreducesthe tioncontrol(Afanasyevetal.,2010). abilityofTCPtofullyutilizetheresourcesofthesenet- works. Duetothisproblemofunder-utilizationofnet- workresources,manyhigh-speedTCPvariantsthataim and how they are behaving. In this paper, high-speed to increasethe utilization ofthese resourceshavebeen LinuxTCPvariantsthatareavailableforresearchispre- exist. These increase of TCP aggressiveness, in order sented and explained, as shown in Table 1, along the tofullyutilizethehigh-speedbandwidths,arisesthese- followingsubsections. vereproblemofburstloss(HaandRhee,2008). Inad- ditiontothat,thevarietyoftheseTCPprotocolsleadsto 2.1. TCPNewReno somequestionsthatneedto beaddressed: WhichTCP variant seems to be the best for high-speed networks? TCPNewRenoisamodificationofTCPRenowhich Are the current TCP variants sufficient to fully utilize developed by FloydandHenderson (1999) then modi- the high-speed bandwidths? In order to answer these fied by Floydetal. (2004), Hendersonetal. (2012) to questions,acomparativestudyofhigh-speedTCPvari- overcomethe problemof Reno’s FastRecoveryduring antsisrequired.Suchcomparisonorperformanceeval- the occurrenceofmultiple packetlosses which signifi- uationaddressesthepointsofTCPweaknessesandcon- cantlydecreasestheReno’sperformanceinheavycon- sequently supports the process of enhancing TCP per- gested networks. In NewReno, the exit from the state formance. of FastRecovery is only allowed if all the data from Nowadays, TCP is struggling to deal with different the initial congestion window are being acknowledged network environments such as wireless or lossy net- which senses the partial data ACKs and differentiates works, high-speednetworksand highlycongested net- itfromnewdataACKs. Morespecifically,thenewdata works. Each type of these networkshas its own prob- ACKreceptionindicatestodeliverysuccessofalldata lems and limitations that are different from one to an- which sent before the loss detection while the partial other networks. Consequently, there are many TCP ACK indicates to other losses in the initial congestion variantsdesignedforeachcertaintypeofnetworks. As window. In fact, NewReno is not designed for high- shown in Figure 1, Afanasyevetal. (2010) provided speed networks (Afanasyevetal., 2010), as shown in an excellent evolutionary graph of most TCP variants Figure1,soitisusedheretobecomparedwiththehigh- basedontheproblemofwhichtheyaretryingtosolve speedTCPvariantsasabenchmark. 2 Table1:TheevolutionofHigh-speedTCPVariantsandtheirimplementationsinthecommonoperatingsystems(Afanasyevetal.,2010). TCPVariant Year Base Windows Linux SunSolaris NewReno 1999 Reno NA >2.1.36 NA HS-TCP 2003 NewReno NA >2.6.13 NA S-TCP 2003 NewReno NA >2.6.13 NA H-TCP 2004 NewReno NA >2.6.13 NA BIC-TCP 2004 HS-TCP NA >2.6.12 NA TCP-AFRICA 2005 HS-TCP,Vegas NA NA NA TCP-Compound 2006 HS-TCP,Vegas XP,Vista,Win7 >2.6.14 NA TCP-illinois 2006 NewReno,DUAL NA >2.6.22 NA TCP-FUSION 2007 Westwood,Vegas NA NA 10,11 YeAH-TCP 2007 STCP,Vegas NA >2.6.22 NA TCP-CUBIC 2008 BIC-TCP NA >2.6.16 NA 2.2. ScalableTCP(STCP) proposed to overcome the poor performance of stan- dard TCP over high-speed networks. HS-TCP is con- STCP was presented at CERN by Kelly (2003) to sidered as loss-based congestion controlalgorithm. In overcomethepoorperformanceoftheexistingconges- fact, HS-TCP did not change the behavior of standard tion control algorithms (such as NewReno) after the TCPthereforeitdidnotpresentanyrisksuchasconges- increase of bandwidths in high-speed networks. The tioncollapse. HS-TCPismerelysender-sidemodifica- challenge for this protocol was to achieve better net- tionwhichincreasesanddecreasescongestionwindow workutilizationwithhigherBandwidthDelayProducts byα(w)andβ(w),respectively. Theresultingfunctions (BDP) withoutcausing any negativeimpacton the ex- α(w)andβ(w)varyfrom1and0.5,respectively,(when istingtraffic.Indeed,STCPismerelysender-sidemodi- thecongestionwindowisbeloworequalto38packet) ficationtotheTCPcongestioncontrolalgorithm.STCP to 70 and0.1, (whenthe congestionwindowis greater hasbeenimplementedinLinuxandthenithasprovided than84kpackets)(Afanasyevetal.,2010,LarandLiao, animprovedperformanceoverthegigabittransatlantic 2013). networkusingstandardTCPreceivers.Atthattime,the Although, HS-TCP succeeded to increase the resultsrevealedthat,theuseofSTCPwouldhaveatriv- throughputin high-speednetworks,itpresentedanag- ialeffectonexistingnetworktrafficatthesametimeas gressive behavior than standard TCP which affects its enhancingdatatransferperformanceinhigh-speednet- sharing fairness especially when competing with stan- works(Kelly,2003). dardTCP flows. Moreover,high-speedTCP presented The loss-based STCP congestion control algorithm another problem over high-BDP networks. This prob- usesα,βwhile(0<α<1)and(0<β<1). STCPup- lem is known as bursty packet losses which is caused dates its congestionwindow after receivingeach ACK bythestandardSlowStartduringthephaseofaninitial in a round trip time by α, as shown in Equation (1), SlowStartwhenanapproximatenetworkcapacityisnot inwhichcongestionisnotdetectedbutifcongestionis yetdetermined.Inordertoovercomethisproblem,HS- detected, it decreases the congestion window by β, as TCPlimitsitsSlowStartto100packets. Thisbehavior showninEquation(2)(Kelly,2003). iswellknownas“LimitedSlowStart”whichisoneof HS-TCPweaknesses. cwnd=cwnd+α (1) cwnd=cwnd−(β∗cwnd) (2) 2.4. HamiltonTCP(H-TCP) H-TCP was presented by D.Leith (2004) at Hamil- Whileαandβsetto0.01and0.125,respectively. tonInstitute. H-TCPisaloss-basedcongestioncontrol protocol,whichissuitableforhigh-speedandlongdis- 2.3. High-speedTCP(HS-TCP) tance networks. It is designed to be more fair and ef- Floyd (2003) proposed a new high-speed TCP for fective than conventional TCP. H-TCP defines the in- large congestion window sizes. This TCP variant was crease in the congestion window w as α(∆) for each 3 RTT(whichincreasesbyafractionα(∆)/wforeachre- were appearing in high-BDP networks. The ag- ceptionofnon-duplicateACK),while∆iselapsedtime gressiveness and scalability of HS-TCP (in case of since last congestion signal. The final function of the congestion-free)and the conservativeattribute of stan- increaseisdefinedasinEquation(3)(Afanasyevetal., dardNewReno(incaseofcongestion)havebeencom- 2010,LarandLiao,2013). binedtogainabetterperformancethantheexistingTCP variants. TCP-Africa is loss-delay-based algorithm, α(∆)=1+10(∆−∆low)+0.5∗(∆−∆low)2 (3) which has borrowed its behavior (congestion/non − congestion)fromTCP Vegasalgorithm; by comparing Where ∆low is a predefinedvalue, whenever∆ < ∆low, the estimated buffer of the network ∆ to a predefined α(∆) = 1. H-TCP reduces its congestion window by constantα. In TCP-Africa, when (∆ < α) which indi- RTTratioEq.(4)ifγEq.(5)islessthan0.2. catestoalittlebufferingspace,itswitchesto fastmode andimmediatelyappliestheCongestionAvoidanceand RTT RTT = min (4) Fast Recovery of HS-TCP algorithm. In this case, the ratio RTT max decrease and increase steps are calculated as β(w) and B(k)−B(k−1) α(w),respectively. Otherwise,itswitchesto slowmode γ= (5) which applies the rules of NewReno that increases by (cid:12) B(k−1) (cid:12) (cid:12)(cid:12) (cid:12)(cid:12) oneaftereveryACKreceptionanddecreasesbyhalving (cid:12) (cid:12) WhereB((cid:12)k)istheestimatio(cid:12)nofachievedthroughputand thecongestionwindowafterlossdetection. B(k−1)istheestimationofprecedinglossevent;oth- TCP-Africa has been evaluated by simulation and erwiseitwillhalveitscongestionwindow. presentedgoodbandwidthutilizationinhigh-BDPnet- works (Kingetal., 2005, Afanasyevetal., 2010). It 2.5. BIC-TCP showedalowerlossratiothanHS-TCPandSTC.Italso presented high fairness (RTT-, intra-, inter-) similar to BIC-TCP was presented by Xuetal. (2004), after thatpresentedbyNewReno.Indespiteofthatimprove- they had pointed out the problem of RTT-unfairness ment,TCP-Africahasnotbeenimplementedinrealop- in HS-TCP and STCP. More specifically, assume that eratingsystems,whereasasimilarmultiple−modecon- twoTCPflowsaresharingonebottleneckandtheyde- gestion algorithm which is Compound TCP has been tect the loss synchronously, if the two flows are HS- implementedin MicrosoftWindowsoperatingsystems TCP, the flow that its RTT is x times smaller can (Afanasyevetal.,2010,LarandLiao,2013). have a network share of x4.56 times larger. But if two STCP flows are used, the smaller RTT will grab all 2.7. TCP-illinois the network bandwidth while the higher RTT will get TCP-illinois was introduced at UIUC by Liuetal. nothing. Hence, BIC-TCP was presented to solve this (2008). It is a sender-side protocol which modifies problem of absolute RTT-unfairness (Harfoush, 2004, AIMDalgorithmofthestandardTCP(Reno,NewReno Afanasyevetal.,2010). or Sack). It uses loss and delay as congestion signals Despite of the improved performance of BIC-TCP, to increase or decrease its congestion window. TCP- its function of window growth can be highly aggres- illinois achieves better performance than the standard sive especially over low-speed or short-distance net- TCPandsharesthenetworkbandwidthfairlyespecially works.Furthermore,BIC-TCPmayachieveabadinter- overhigh-BDPnetworks. TCP-illinoisupdatesitscon- fairness and RTT-fairness due to its dependability on gestion window after every ACK reception in a round RTTmeasurements. Aswellas,ithasahighcomplex- trip time by(α/cwnd)in whichcongestionhas notde- ity due to the several modes (binary search increase, tected but when congestion detected, TCP-illinois de- max probing, Smax and Smin) of the algorithm itself. creasesitscongestionwindowby(β∗cwnd)asinEqua- Thus,BIC-TCPhasbeenreviewedandmodifiedinCU- tions(6)and(7)(Liuetal.,2008),respectively. BIC which conserves the stability and scalability of cwnd =cwnd+(α/cwnd) (6) BIC-TCP, decreases the complexity, and increases the fairness(HaandRhee,2008,Afanasyevetal.,2010). cwnd =cwnd−(β∗cwnd) (7) TCP-illinoisuses loss signalto set the directionand 2.6. TCPAfrica use delay to calculate the step of window size change TCP-Africa(AdaptiveandFairRapidIncreaseCon- by f (.) and f (.) as explained in reference (Liuetal., 1 2 gestionAvoidance)waspresentedbyKingetal.(2005). 2008), while (0 ≤ α ≤ 1), (0.125 ≤ β ≤ 0.5)andα = TCP-Africa was designed to solve the problems that f (d ),β= f (d ),where(d )isdelay-average. 1 a 2 a a 4 2.8. CompoundTCP(C-TCP) 2.10. TCPFusion Kanekoetal. (2007) presented TCP Fusion which TanandSong (2006) introduced new loss-delay- combining Westwood′s achievable rate, DUALs based TCP variant named C-TCP. As TCP-Africa, C- queuing delay, and Vegas used network buffering TCPcombinestwomodesofNewRenoandHS-TCPto estimations. Dependingontheabsolutethresholdvalue increase the bandwidth utilization over high-BDP net- ofqueuingdelay,Fusionswitchestoitsthreemodes;if works. C-TCP compares α to the estimated ∆, where thequeuingdelayislowerthanthepredefinedthreshold, αis smallpredefinedconstant. When∆ exceedsα, C- the fastmode is appliedwhichincreasesitscwnd bya TCPgentlyreducesW byapredefinedζ asshownin fast predefinedachievablerate estimationfractionofWest- Equation(8)(Afanasyevetal.,2010). wood.Whileifthecurrentqueuingdelayisgreaterthan three times of the threshold, cwnd is decreased by the W =W −(ζ∗∆) (8) fast fast numberofbufferedpacketsinthenetwork. Otherwise, if the queuing delay is somewhere between one and C-TCPcalculatesW toaddittothefinalconges- fast three times of the predefined threshold, Fusion keeps tionwindowasshowninEquation(9)(Afanasyevetal., its cwnd as it is. Indeed, experimentalresults showed 2010). theimprovementofFusionperformancemetricssuchas bandwidthutilizationandfairnesscomparedtoC-TCP, W =W +W (9) reno fast HS-TCP, BIC and Fast. Despite of the improvement, Fusionhasmanylimitationssuchastheproblemofpre- This Wfast is a smooth movement from HS-TCP definingthethresholdwhichisdonemanually,andthe fastmode to NewReno slowmode. C-TCP behavioris more critical problemwhich may lead Fusion in some verysimilartoTCP-AfricabutC-TCPshowsaconvex cases to behave similar to standard NewReno most of curveafter exceedingthe thresholdwhile TCP- Africa thetime(Afanasyevetal.,2010). shows linear increase. In despite of the changes in its behavior,C-TCPstillachievesassameperformanceas 2.11. CUBICTCP TCP-AfricaandevenpresentsanotherproblemofRTT CUBIC TCP was presented by HaandRhee (2008) estimation which is inherited from TCP Vegas. This antitisthecurrentdefaultTCPalgorithminmostLinux problemmakesC-TCPverysensitivetoRTTmeasure- operating systems. It modified the linear function of mentswhichmakesitslightlyunfair. However,C-TCP cwnd increase in the existing TCP variants to cubic iscurrentlythemostdeployedcongestioncontrolalgo- function in order to enhance its scalability over high- rithm since its implementation in Microsoft Windows BDPnetworks.HaandRhee(2008)havereviewedBIC operatingsystems(Afanasyevetal.,2010). algorithm to come up with CUBIC which borrowed the cubic function of congestion window from H-TCP 2.9. YeAHTCP approach as shown in Equation (10) (Afanasyevetal., 2010). YeAH(YetAnotherHigh-speed)TCPwaspresented 3 byBaiocchietal.(2007). ItissimilarinspiritofTCP- β∗w w=C ∆− 3 max +w (10) AfricaandC-TCP.ItcombineslossdetectionandRTT r C max ecsotmimbaintieosnNteowpRreednioctanndetSwToCrkPdineslateya.dSoifmHilSa-rTlyC,PY,esAoHit where C isa predefined constant, β is a coefficient of increasesthecongestionwindowbyoneeveryRTTand multiplicativedecreaseinFastRecovery,andw isthe max halving it if a loss is detected (by receiving three du- congestion window size just before the last registered plicated ACKs). More specifically, if (∆ < α), where lossdetection.LimitedSlowStart,RapidConvergence α is a predefined threshold, and (Q/RTT < φ), andRTT independenceinCUBIC,allprovidedhigher min whereφisanotherpredefinedthreshold,YeAHswitches fairness(RTT-,intra-)andhigherscalability. Thetarget to fastmode and behaves similarly as STCP. Other- windoww iscalculatedintheinitialstageofthewin- max wise, a slowmode of NewReno is applied. Briefly, dowincreasewhichisdiscoveredbytherightbranchof YeAH showed higher efficiency and fairness (inter-, cubic function. The exponential increase of standard intra-, RTT-) than TCP-Africa and C-TCP especially SlowStartismoreaggressivethanthediscoveryofthe in high-BDP networks but it still has the same prob- window increase which is more scalable in high-BDP lem of RTT estimation which is inherited from Vegas networks. Upon loss detection, if this loss is tempo- (Afanasyevetal.,2010,LarandLiao,2013). raryandw isnotreachedyet,cwndwillbeincreased max 5 accordingto both right and left branchesof the cubic throughput over high-BDP networks. In order to in- function. crease the bandwidth utilization, Khalil (2012) pro- Moreover,CUBICensuresthat,itsthroughputisnot posedanewcongestioncontrolschemecalledSwift− lower than the throughput of the standard NewReno, Start. It changes the way of estimating the available whichisdonebyenforcingthecalculatedvalueofw bandwidthtoavoidthecongestionwhichcausedbyover reno whenever w is going lower than w . This com- orunderbandwidthestimation. Cavendishetal.(2012) max reno plicatedbehaviorofCUBICalgorithmconfirmsavery prove that TCP can achieve a superior performance if highperformanceandfairnessattributes,whichmakeit itsparametersaretunedwelldependingonnetworkand thesecondmostusedTCPvariantafterbeingthestan- pathconditions. dard TCP of Linux operating systems. However, CU- Moreover, network buffers are going to- BIC is still have some limitations that lead to under- wards the near − zero buffer, as mentioned in utilization of the available bandwidth and produces a (Enachescuetal., 2006, Beheshtietal., 2006, huge number of packet losses especially in high-BDP Prasadetal., 2007, VishwanathandSivaraman, 2008, networks. Theselimitationsareduetothedependency Vishwanathetal.,2009a,b,VishwanathandSivaraman, oflosswhichistheonlycongestionsignalusedinthis 2009, Sivaramanetal., 2009, LeGrangeetal., 2009, algorithm (Afanasyevetal., 2010, LarandLiao, 2013, Vishwanathetal., 2011), to fit the all-fiber networks HaandRhee,2008). which is the fastest type of high-speed networks yet. Consequently, it is very important to take the case of 2.12. LatestIssues near − zero buffer network into account in the future Fuetal.(2007),MohamedA.AlrshahandMohamedOthmTaCnPperformanceevaluation. (2009), Qureshietal. (2012) and Asmentionedabove,TCPisstillsufferingfrommany MohamedA.AlrshahandMohamedOthman (2013) problems, and researchers are still modifying and im- confirmed that, the single-based TCP with an appro- provingit. Someresearchersmixdifferentmodes,such priate modification can overcome and well replace asfastandslowmodes,andswitchbetweenthembased the parallel-based TCP and it may be able to fully on the state of the network. And some researchers utilizethehigh-speedbandwidths. WhileHaandRhee mixdifferentapproaches,suchasloss anddelaybased (2011) mentioned that, standard Slow Start becomes approaches, to improve the performance. And some inappropriate for the high-BDP networks and they researchers are estimating the RTT and bandwidth to statedtworeasonsforthisproblemasbelow: avoid the severe congestionwhich can lead to conges- tion collapse. While some are tryingto modifythe al- 1. The exponential increase of the congestion win- gorithm itself by modifying Slow Start or Congestion dow results a heavy packet losses that make the Avoidancealgorithm,andsomeoftherestaretryingto entire system completely unresponsive for a long tunetheTCPparameterscarefullytoachieveasuperior periodoftimeduringthelossrecoverystage. performance. 2. Some optimizations, that applied to Slow Start, happen to slow down the loss recovery followed by Slow Start which leads to under-utilization of 3. PerformanceEvaluationofTCPVariants thenetworkresources. Inordertosolvetheabovementionedproblems,they Inthispaper,twosimulation-basedexperimentshave presented a new Slow Start algorithm named ”HyS- been conducted to show the performance differences tart”. This algorithmfinds a safe exitpointfromSlow among high-speedLinux TCP variants overhigh-BDP StarttoCongestionAvoidancewithoutcausingaheavy congestedandnon-congestednetworks.Thefirstexper- packetlosses. Thisalgorithmimprovesthethroughput iment has been conductedto evaluate the performance ofTCPandithasbeenalreadyappliedtoCUBICsince ofTCPovernon-congestednetworktomimictheideal Linux 2.6.29 as a default Slow Start. Xuetal. (2011) case of the network, then, to show the ability of TCP proposedanewhybridcongestioncontrolcalledHCC- onbandwidthutilization,andtodeterminethepointsof TCP which is loss-delay-based. HCC-TCP improves weaknesses in its mechanism. In addition to that, the thethroughputandfairnessaswell. second experimenthas been conductedto evaluate the DangiandShukla (2012a) proposed a new hybrid performanceofTCPovercongestedbottleneckinorder (loss-delay-based)congestioncontrolscheme. Theex- tosimulatearealnetworkscenario. periments of DangiandShukla (2012b) reveals that, InthefirstexperimentTCPvariantshavebeenevalu- HCC-TCP can achieve an efficient performance on atedbymeasuringtheaveragethroughputandlossratio 6 while in the second experiment they have been evalu- Table2:ExperimentParameters. ated by measuring the average throughput, loss ratio, intra-fairnessandRTT-fairness.Morespecifically,mea- suringtheaveragethroughputisbeneficialtoshowthe No.Parameter Value abilityoflinkutilization,whilemeasuringthelossratio 1. TCPVariants Scalable, HS-TCP, BIC, H- is helpfulto show the quantityof lost data which neg- TCP,CUBIC,Africa,C-TCP, ativelyaffectsthe generalperformanceofTCP. On the Fusion, NewReno, illinois otherhand,measuring(intra-,RTT-)fairnessistoshow andYeAH. the quality of sharing the link between the competing 2. Linkcapacity 1000Mbpsforall. TCP flows based on Jain’s fairness index (Jainetal., 3. Linkdelay 1msnodetorouter. 1984). All of these measurements are conducted to 100msroutertorouter. show the advantagesand disadvantagesof all involved 4. BDP 12750KB (High-BDP as TCP variants to determine the points of strengths and in JacobsonandBraden weaknesses of every TCP variant in order to help the (1988)). processofimprovingtheperformanceofthesevariants. 5. Buffersize from100to5000packets. 6. Packetsize 1000bytes. 3.1. ExperimentsSetup 7. QueuingAlgo DropTail. 8. Traffictype FTP In the first experiment, a standard single dumbbell 9. Simulationtime 100seconds. topologyhasbeenusedasshowninFigure2. Onlyone sender (S1) and one receiver (D1) are used. S1 sends data to D1 through two routers on the path. S1 and D1areconnectedtotheroutersoverLANwith1Gbps speed and 1ms propagation delay. While the routers 100 to 5000 packets. In fact, this experiments show arelinkedby1Gbpsspeedwithapropagationdelayof the impact of bottleneck congestionand buffer size on 100ms.Asitisclearinthistopology,thisnetworkdoes theperformanceoftheexaminedTCPvariantsandalso nothavebottleneck,thereforeitisconsideredasthebest show the performance changes when a smaller buffer caseofTCPoveranidealnetwork. size is applied. Scalable, HS-TCP, BIC, H-TCP, CU- BIC,Africa,Compound,Fusion,NewReno,illinoisand YeAH are involved in these experiments. All of these TCPvariantsareaddedintoNS2version2.35whichis installedonLinuxopenSuse12.2,kernelversion3.4.28 overIntelCore-i7machineto performthis simulation- basedcomparison. Table2showstheexperimentsetup Figure2:Non-congestednetworktopology. andthesimulationparameters. As for the second experimentsetup, a standard sin- gledumbbelltopologyhasbeenusedas showninFig- ure 3. As shown in the network topology, there are n competingsenders(S1,S2, S3,..., Sn)senddatasyn- chronouslytonreceivers(D1,D2,D3,...,Dn)througha sharedsinglebottleneck. Allnodesofsourcesanddes- tinationsareconnectedtobottleneckroutersoverLAN with 1Gbps speed and 1ms propagation delay. While the bottleneck link is 1Gbps speed with a propagation delay of 100ms. Consequently, the proper bandwidth of the shared bottleneck, which is needed by the con- current senders, is 4Gbps while the available is only 1Gbps, this in order to simulate a real congested bot- tleneck. Figure3:Networktopologywithstandarddumbbellbottleneck. This experiment is repeated for every TCP variant separatelywith differentbuffer sizes which starts from 7 3.2. ResultsandDiscussion thefairnessbyreducingthegabbetweenthemaxlimit ofthebandwidthandthereducedcwndafterlossdetec- Based on the results of the first experiment, it is tion. brieflyconcludedthat,TCPSlowStarthasafatalprob- lemknownasburstloss. BurstlosshappenswhenTCP Asfortheresultsofthesecondexperiment,Figure5 jumpsexponentiallytoreachthemaximumcwndinor- shows that, CUBIC, BIC and YeAH are the best three dertoquicklyutilizethebandwidthofthenetwork.The TCPvariantsintermsofthroughput.Morespecifically, lackoftheinformationaboutthelinkbandwidthmakes inthecasesof5000,2500and1000packetsbuffersize TCPincreasesitscwnduntilitdetectsthefirstlossthen YeAHisabitbetterthanCUBICandBIC.Whileinthe ithalvesitscwndandentersthelinearincreasestage. In other cases of smaller buffer sizes, CUBIC overcomes otherwords,theburstlosshappenswhenthefirstlossis all other TCP variants whenever buffer size is going detected. Indeed,thisburstlosscanseverelyaffectthe closertonear−zerobuffer.Asmentionedabove,YeAH, performanceofTCPanditmayevenleadtocongestion BICandCUBICareusingaconservativeandgentlere- collapse. duction of cwnd which helps them to outperform the All the TCP variants in this experiment, jump to other variants of TCP that half their cwnd whenever a reachacwndofaround60000packetsbutafterthefirst lossisdetected. lossisdetected,theirbehaviorsbecomecompletelydif- Indeed, average throughput solely is not enough to ferent. NewReno, Africa, illinois, C-TCP and Fusion evaluate the performance of TCP variants. The other droptheircwnd tothehalfthenincreaseitlinearlyina important performance metric, which is necessary to veryslow mannerasshowninFigures4(a), 4(b), 4(c), evaluate the performance of TCP, is the loss ratio. In 4(d)and4(e), respectively. Thislinearincreasebehav- Figure6,thelowestlossratioisNewReno,whichhasa ior consumes a long time to reach the upper limit of verypoorthroughputaveragebecauseitisnotdesigned the network bandwidth again which results an under- for high-speednetworks. For thisreason it will notbe utilizationofthenetworkresources. considered. Unlikely,scalableandcompoundhavethe Diversely, STCP, HS-TCP, H-TCP and YeAH drop highestlossratio. While,BIC,CUBICandH-TCPare their cwnd to the half then increase it in an oscillating keepingtheirlossratiostableregardlessofthechanges manner as shown in Figures 4(f), 4(g), 4(h), 4(i), re- inbuffersize. Infact,allofthecomparedTCPvariants spectively. This behavior increases the bandwidth uti- produce similar number of lost packets, but when the lization to some extent, but some of these TCP vari- countoflostpacketscomparedtothetotalsentdataas ants such as STCP, HS-TCP and H-TCP are still suf- alossratio,itwillgiveacompletelydifferentreadouts. fer from the problem of under-utilization while YeAH hasachievedhigherbandwidthutilizationduetoitscon- As regarding to (Intra-, RTT-) fairness as shown servativereductionintheCongestionAvoidancestage. in Figures 7 and 8, Scalable, Africa, Fusion, high- More specifically, all of the aforementionedTCP vari- speed, YeAH and NewReno are oscillating and they ants reduce their cwnd to the half when the loss is de- showdifferentfairnessindexwheneverthebuffersizeis tected, butin YeAH,cwnd is reducedto the half if the changed. On the other hand, Compound, H-TCP, CU- loss is detected in Slow Start stage while it reduced BIC, BIC and illinois are the best and they are mostly gentlyintheCongestionAvoidancestagebasedonthe very close to 1 which represents the best case of fair- changesofnetworkdelay. ness.AsillinoisandCompoundpresentsapooraverage More differently, BIC and CUBIC drop their cwnd, throughputandasCUBICisaderivedversionfromBIC thenincreaseitinarapidconvergencemannerasshown andH-TCPsothebestTCPvariantintermsofaverage in Figures 4(j) and 4(k), respectively. Indeed, they re- throughput, loss ratio and fairness is CUBIC followed ducetheircwnd toaround85%ofthelastcwnd when- byYeAH. ever a loss is detected regardlessof the loss is coming fromSlowStartorCongestionAvoidance. Thisbehav- As described in Section 2, CUBIC is a loss-based ior results in a higher bandwidth utilization than the algorithm and YeAH is a loss-delay-based algorithm. other TCP variants which halving their cwnd after ev- Thus, both approaches can give a higher performance ery loss. As a final point, and based on observation, than the otherexisting TCP variantsandboth can pro- the conservative reduction of the cwnd can help TCP vide similar performance in the case of high-speed toincrease: (1)thebandwidthutilizationbyreleasinga wired networks. Despite of all, the bandwidth utiliza- smallpartoftheusedbandwidth,aftereveryloss,which tion of CUBIC and YeAH, is still not enough to cope helpstoreachagaintothehigherlimitmorefaster. (2) withthenewgenerationofthesenetworks. 8 60000 60000 60000 newreno africa illinois 50000 50000 50000 WND (Packets) 3400000000 WND (Packets) 3400000000 WND (Packets) 3400000000 C 20000 C 20000 C 20000 10000 10000 10000 0 0 0 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 Simulation Time (Seconds) Simulation Time (Seconds) Simulation Time (Seconds) (a) TCPNewReno (b) TCPAfrica (c) TCPillinois 60000 60000 60000 compound fusion scalable 50000 50000 50000 WND (Packets) 3400000000 WND (Packets) 3400000000 WND (Packets) 3400000000 C 20000 C 20000 C 20000 10000 10000 10000 0 0 0 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 Simulation Time (Seconds) Simulation Time (Seconds) Simulation Time (Seconds) (d) CompoundTCP (e) TCPFusion (f) ScalableTCP 60000 60000 60000 highspeed htcp yeah 50000 50000 50000 WND (Packets) 3400000000 WND (Packets) 3400000000 WND (Packets) 3400000000 C 20000 C 20000 C 20000 10000 10000 10000 0 0 0 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 Simulation Time (Seconds) Simulation Time (Seconds) Simulation Time (Seconds) (g) High-speedTCP (h) HamiltonTCP (i) YeAHTCP 60000 60000 bic cubic 50000 50000 WND (Packets) 3400000000 WND (Packets) 3400000000 C 20000 C 20000 10000 10000 0 0 0 20 40 60 80 100 0 20 40 60 80 100 Simulation Time (Seconds) Simulation Time (Seconds) (j) BicTCP (k) CubicTCP Figure4:ThecongestionwindowdynamicsoftheexaminedTCPvariants. 4. Conclusion loss can dangerously congest the bottleneck and may leadtoslowdowntheperformanceofTCPandevento In a nutshell, TCP Slow Start has a fatal problem congestioncollapse. AlltheaforementionedTCPvari- whichknownastheburstlosswhichhappensattheend antssufferfromthisproblemindifferentlevelsofdan- oftheinitialstageofSlowStart. Thecauseofburstloss ger. istheexponentialincreaseofthecongestionwindowin ordertoquicklyutilizethebandwidth.Indeed,theburst AsshowninFigure5,itisveryclearthat,thesmaller 9 scalable illinois yeah cubic scalable illinois yeah cubic compound fusion htcp bic compound fusion htcp bic highspeed africa newreno highspeed africa newreno 1000 900 1 ps) 800 0.998 erage Throughput (Mb 345670000000000 RTT-fairness (Index) 000...999999246 Av 200 0.99 100 0 0.988 100 250 500 1000 2500 5000 100 250 500 1000 2500 5000 Buffer Size (Packets) Buffer Size (Packets) Figure5:AverageThroughputvs.BufferSize Figure8:RTT-fairnessvs.BufferSize scalable illinois yeah cubic compound fusion htcp bic highspeed africa newreno the available TCP variantshave presetvariableswhich makeitmorestaticandneedsadifferentsettingforeach 0.45 0.4 network scenario. The existence of these preset vari- 0.35 ables makes the implementationof these TCP variants 0.3 moreharderandreducesitsadaptationcapabilitiestofit %) o ( 0.25 differentscenarioswithoutanymanualchanges. Thus, Lost Rati 0 .01.52 isnetfvuaturiraebvleesrssihoonuslodfbTeCaPvopidroedtoicnolo,rtdheerutoseinocfrethaeseptrhee- 0.1 adaptationcapabilitiesoftheprotocol. 0.05 Furthermore, CUBIC (loss-based) is the best TCP 0 variantwhichovercomesallothervariantsinmostsce- 100 250 500 1000 2500 5000 Buffer Size (Packets) narios. YeAH(loss-delay-based)showsa goodperfor- manceinmostcasesbutthey(CUBICandYeAH)still Figure6:LossRatiovs.BufferSize produceahugeburstloss. Consequently,thisburstloss may be avoided in the future, by implementing a way scalable illinois yeah cubic compound fusion htcp bic which has the ability to find a safe exit pointfrom the highspeed africa newreno SlowStartphasebeforetheoccurrenceoftheburstloss. 1.05 This safe exit point may be found by estimating the 1 available bandwidth or by calculating the chain of the ex) 0.95 ACKarrivals. nd 0.9 For more details, Table 3 shows the results of the ess (I 0.85 second experimentin detail. While Throughputis the n a-fair 0.8 rateofsuccessfullyreceivedpacketsmeasuredasMbps. Intr 0.75 LossRatio refers to the ratio between the total number of lost data packets to the total of sent packets. Intra- 0.7 fairandRTT-fairdeterminewhethertheconcurrentTCP 0.65 100 250 500 1000 2500 5000 flows are receiving a fair share of network bandwidth Buffer Size (Packets) andtime,respectively. Intra-fairandRTT-fairaremea- Figure7:Intra-fairnessvs.BufferSize suredasindexfromzerotoone,whilethehigherindex isthehigherfairness. buffersizeisthelowerTCPthroughput.Thus,allofthe existingTCPvariantsstillrequiremoreimprovementto Acknowledgment extendtheirabilitytofullyutilizethehigh-speedband- widths,especiallywhentheappliedbufferisnear−zero This work was supportedby the Ministry of Higher or less than the BDP of the link. Furthermore, all EducationofMalaysiaundertheFundamentalResearch 10