roceedngs of the 7th World Congress The Internatonal Federaton of Automatc Control A assve Network Measurement-based Traffc Control Algorthm n Gateway of 2 Systems Ybo Jang, Weje Chen, Janwe Zheng, Wanlang Wang*, Yanwe Zhao Collage of Informaton Engneerng, Zhejang Unversty of Technology,.R.Chna (*Correspondng author, e-mal: wangwanlang@zjut.edu.cn) Abstract: Wth the progress of peer-to-peer technology, 2 applcatons have evolved and establshed themselves as the leadng growth app of Internet traffc workload. In order to reserve enough bandwdth for other applcatons, gateway of 2 systems must control 2's traffc. In ths paper, passve network measurement of 2 traffc s ntroduced, and network behavors of some 2 applcatons are dscussed. Then, one model of recevng and forwardng 2 packets s constructed. One algorthm of traffc control n 2 s desgned, whch ncludes two mportant factors: rejecton probablty of 2 request and dsconnecton probablty of 2 channel. The results of two parallel experments show that control algorthm of 2 traffc based on network measurement can reserve enough network resource wthout dsablng all 2 packets.. INTRODUCTION From the end of 9th century to the begnnng of 20th century, 2 (peer-to-peer) technology has been greatly mproved. Applcatons of 2 systems nclude fle sharng systems such as Napster (Ben et al., 2000), bt-torrent (Bram Cohen, 2007), edonkey (Olver et al., 2002), Kazaa, Gnutella, BtComet as well as VoI systems such as Skype and GoogleTalk. Recently, some tradtonal dstrbuted applcatons begn to use 2 technology, such as GRID, ITV, networked games, etc. 2 systems are so popular that they contrbute more than ffty percent to the overall network traffc. Thus, 2 systems have ganed a lot of attenton from the Internet users, ISs (Internet Servce provders) and research communtes. 2 technology s one of dstrbuted applcatons, whch does not depend on network server. In other words, there s no center node n 2 systems, and 2 s network s organzed by ts self freely. In such stuaton, 2 technology can solve the problem of bottleneck of server n data transmsson of clent/server mode (Andrew et al., 200). 2 technology makes resource sharng convenently. However, the wde-spread use of 2 systems has brought some serous drawbacks to Internet. To maxmze speed of resource sharng and transmttablty of data, many 2 softwares try to use as much bandwdth as they can. In practce, 2 software often creates more than ten threads smultanety to send and receve data, and search data source n other nodes contnually at the same tme. 2 software s actons make network much more congested than before (Savage et al., 999). ISs have been put n a dlemma. On the one hand, 2 system applcatons have resulted n an ncrease n revenue for ISs, as they are one of the major reasons cted by Internet users for upgradng ther Internet access to broadband. On the other hand, ISs fnd that 2 traffc poses a sgnfcant traffc engneerng challenge (ablo et al., 2006) (Vnay et al., 2007). 2 traffc always occupes other applcatons bandwdth, such as Web, FT and Emal. In local area networks, people are accustomed to use 2 softwares to share fles, play games and exchange moves. Not only 2 systems wll acqure much network resource n Internet, but also t wll lead to congeston n ntranet. Because 2 network s a random overlay network whch makes ntranet as a part of Internet, 2 systems network usually through gateway between Internet and ntranet. In recent years, some ISs have prepared to control traffc of 2 applcatons and many network managers of ntranet dsabled 2 applcaton drectly. Obvously, data sharng s the basc demand of network users, and 2 systems can make the speed of data transmttng faster than any other tradtonal applcatons. Although 2 systems have some drawbacks n traffc, t s especally rratonal that we dsable 2 applcatons for no reason other than ths smple fact. Furthermore, some 2 systems have modfed ther protocols and softwares to hde ports of TC protocol and payload of I packet, today. In our opnon, 2 technology can be greatly developed and became the frst applcatons of Internet reveals that t s a very effcent technology n many areas. It s not a good dea to avod network congeston by smply dsable 2 978--234-7890-2/08/$20.00 2008 IFAC 209 0.382/20080706-5-KR-00.2782
applcatons. So, we need to fnd an easy way to control 2 traffc. We wll ntroduce network measurement and some mportant network parameters n 2 systems frst. Then, some dfferent methods of measurement wll be classfed. And we analyze common behavors of 2 softwares based on network measurement and fnd a relatonshp between 2 traffc and behavor. At last, we wll construct a model of 2 traffc and desgn an algorthm for traffc control. Am of control algorthm n gateway s to farly assgn bandwdth for 2 applcatons and other applcatons. between source node and destnaton node wll not be constructed. At runtme of 2 softwares, they wll send many request packets to other nodes, and search new neghbor nodes to ncrease transmttng. We have traced the bandwdth of 2 traffc n a local area network, as fgure 2. 2. NETWORK MEASUREMENT AND SOFTWARE BEHAVIOR IN 2 SYSTEMS Man task of network measurement s to measure some parameters of real networks, such as bandwdth, packet delay, drop rate, etc. Then we can evaluate the runtme status of network by these statstc data (Stemm et al., 2000) (Wjata et al., 2000). 2 system s network s one type of overlay networks whch based on Internet and local area networks. Thus, network measurement n 2 systems must collect parameters of 2 s network and nfrastructure of overlay network. There are two types of network measurement n 2 systems, actve measurement and passve measurement. Actve network measurement modfes a 2 software clent manually, and connects ths clent as normal 2 software nto 2 network to detect characterstcs of entre networks. Thus, ths clent s seemed as a common node by other 2 softwares among neghbor nodes. Actve network measurement often collects I address, transportaton port, request port and sharng resources of other nodes, as well as topology structure, packet delay, upload speed and download speed of 2 network. assve network measurement usually deploys some detecton ponts n nfrastructure network. And t uses specal softwares and hardwares to collect nformaton of 2 traffc (Sen et al., 2004). Wth the developng of 2 software and 2 transmttng protocol, 2 applcaton s TC port can be automatcally changed and 2 packet s payload s encrypted. assve network measurement s ablty has been lmted. Actually fact, network measurement n 2 systems s to detect network behavors of softwares whch mplement 2 protocol. Sarou et al. analyzed two 2 systems, Napster and Gnutella (Sarou et al., 2002). ouwelse et al. analyzed BtTorrent whch s the most popular 2 system today (ouwelse et al., 2004). By ther researches, we fnd such a common characterstc of behavors that all types of 2 softwares wll send many request packets to ts neghbor nodes whch are detected before, whle a new node s connectng nto 2 network. And neghbor nodes wll send back response packets whch contan ther local nformaton to new node after they receve a request packet. Although, dfferent types of 2 softwares may use dfferent searchng methods of neghbor node and collecton algorthms of network nformaton, they must send a request packet to other node for handshake before data sharng, as fgure. If request of data sharng s faled, data transmttng Fg. Request and response process of 2 softwares Fg 2. Bandwdth of 2 data sharng n AN Fgure 2 shows that 2 software wll search other neghbor nodes by floodng lke method and fnd out as many resources as possble. Because of random dsappearance and appearance of neghbor nodes n 2 network, 2 software would update neghbor nformaton dynamcally. If we can dsable request packets (aurent et al., 2006), gateway would prevent constructon of 2 data transmttng channel and 2 software s data sharng process wll be controlled. Hence, we should desgn a control algorthm to control 2 traffc and make t works wthout nfluence other applcatons. 3. CONTRO AGORITHM OF 2 TRAFFIC Control algorthm s based on network measurement, and t wll control 2 traffc n gateway s packet flter. The man task of control algorthm s to control and lmt 2 traffc that guarantee other applcatons would assgn enough 2020
network resource farly. There are two control methods. The frst method s lmtaton of 2 request packet dependng on gateway s dynamc load n tme. The second method s dsconnecton of some 2 data transmttng channels whch have been occuped more network resource when gateway s overloaded. Control algorthm needs some parameters of network measurement, such as type dentfcaton of every workstaton s packet, statstc of traffc n delta tme, sendng and recevng rate of packet, request packet dentfcaton. Number of workstatons n a local area network s n, and speed of data transmttng of every workstaton s v ( = n). v s calculated by network measurement module, whch s mplemented n gateway. Network measurement module can record every packet through gateway, and dentfy whch one s 2 data packet. A packet can be dentfed by ts TC port and payload nformaton. By every speed v, current total speed of 2 traffc n gateway s V, as follows: V = n v = Network manager could custom 2 s workload of gateway dependng on real network s status, and range of load s. So bandwdth of 2 traffc should be S [ 0, B] controlled n S. Moreover, ths range s an estmated range. Real bandwdth of 2 traffc may beyond the upper lmt n a short tme. Gateway should decde whch request packet s allowed to forward whle a packet of 2 traffc s receved from workstatons. If a request of 2 system s allowed, t wll be forwarded mmedately. Otherwse, the request packet wll be dropped drectly. All types of 2 softwares have fault-recovery mechansm of tmeout. If a request s tmeout, the data transmttng channel wll not be constructed. Control algorthm n gateway can lmt 2 traffc by droppng request packet. Control algorthm use a random method to reject request packet by rejecton probablty of request α, as follows: V e B ( V ) = tv exp B t () α (2) t s a constant. Network manager can change t to modfy rejecton probablty of request. Every workstaton s 2 traffc n prevous span s accumulated n ( = n), and proportons of total D bandwdth s ( = n), as follows: N N = n D j= Every node s percentage of current total bandwdth of 2 traffc s ( = n), and M = v V. Every node s M D j / dsconnecton probablty of 2 data transmttng channel s =, as follows: ( n) exp( t tn ) exp( t tm ) t (3) t t ( M, N ) = N + wm (4) + w w s the weght rato of by network manager. M and + w N, and t s adjusted Modfed dsconnecton probablty of 2 data transmttng channel for every node s ( = n), as follows: β j= j β = n (5) In ths model, there are two probabltes, rejecton probablty of request α and dsconnecton probablty of 2 data transmttng channel β. Rejecton probablty of request α nfluences packet swtchng process when 2 software begns to connect nto overlay network. By current traffc parameters, gateway randomly drops every request packet dependng on rejecton probablty. Usng random drop method can prevent later nodes have no chance to start a 2 connecton. Dsconnecton probablty of 2 channel β wll be used when gateway s overloaded. Gateway detects total bandwdth of 2 traffc based on network measurement module n runtme. If total bandwdth of 2 traffc beyond the upper lmt of estmated range S, gateway wll randomly nterrupt more than one 2 data channels dependng on dsconnecton probablty. At the same tme, gateway wll reject all requests packet, because rejecton probablty s one. After that, new 2 data channel wll not be constructed and old 2 data channels whch usng more network resource have hgher probablty to be nterrupted. Thus, gateway wll force 2 systems release bandwdth untl gateway s load reduce to S. 202
4. IMEMENT AND EXERIMENTS Accordng to prevous sectons, we constructed a model for 2 software and desgned an algorthm to control 2 traffc. Then, we mplemented ths algorthm as a computer functon n gateway. Control functon s embedded nto packet flter system. And packet flter software of gateway collects every node s parameters, such as packet sendng rate, packet recevng rate and TC port. Network measurement module dentfes every packet by TC port and record packet nformaton of 2 traffc and other applcatons traffc, respectvely. Control functon whch contans control algorthm wll calculate rejecton probablty of request and dsconnecton probablty of 2 data transmttng channel to determne how to control 2 traffc n real-tme. acket flter was deployed n a gateway, and the local area network had fve workstatons. Every workstaton had some dfferent 2 softwares, such as BtComet, Emule and BtTorrent. To accurately dentfy every 2 packet, we bound 2 softwares to fxed TC port. In control algorthm, constant t was 0 and w was 4. 2 estmated bandwdth was S [ 0kB / s,80kb / s]. All 2 softwares started together and shared data wth other neghbor nodes n Internet. Experments last 300 seconds and record bandwdths every 2 seconds. Two parallel experments are run for contrast. The frst experment (Exp.) was the common packet flter model whch was not mplement control algorthm. The second experment (Exp.2) used packet flter wth control algorthm. After experments, contrast charts of bandwdth of fve workstatons as fgure 3 and contrast chart of total bandwdth of gateway as fgure 4. Results of two experments show that f gateway s packet flter dd not use control algorthm, every 2 software n workstatons would search other neghbor nodes and constructed many data channels for transmttng. Total bandwdth of 2 traffc has been ncreased quckly untl gateway overloaded. Usng control algorthm, however, bandwdth of 2 traffc have been controlled well n estmated range. Thus, gateway has enough bandwdth to serve other applcatons. By our algorthm, AN can control 2 systems traffc, and gateway may assgn network resource for every applcaton farly. Fg 4. Contrast chart of total bandwdth. (kb/s) Fg 3. Contrast charts of fve workstatons bandwdth. (kb/s) 5. SUMMARY In order to reserve enough bandwdth for other applcatons, gateway must control 2 s traffc. In ths paper, we ntroduced network measurement and some mportant network parameters n 2 system. Some dfferent methods of measurement were classfed. And we analyzed common behavors of 2 softwares based on network measurement and found a relatonshp between 2 traffc and behavor. Then, we constructed a model of 2 traffc and desgned an algorthm for traffc control. At last, two contrast experments had been performed. The results of two parallel experments show that control algorthm of 2 traffc based 2022
on network measurement can reserve enough network resource wthout dsablng all 2 packets. 6. ACKNOWEDGEMENT Ths work was supported by the Natonal Scence Foundaton of Chna (6057323). REFERENCES mproved performance? [J], ACM SIGCOMM Computer Communcaton Revew, Vol37(3), pp29-40. Wjata Y.I., Nehaus D., Frost V.S. (2000), A scalable agent-based network measurement nfrastructure [J], Communcatons Magazne, IEEE, Vol(38):9, pp74-83. Andrew Oram, Andy Oram (200), eer-to-eer: Harnessng the ower of Dsruptve Technologes [M], O'Relly & Assocates, Inc., Sebastopol, CA. Ben Byer, Evan Martn, Colten Edwards (2000), Napster messages, http://opennap.sourceforge.net/napster.txt. Bram Cohen (2007), Bttorrent protocol specfcaton, http://wk.theory.org/bttorrentspecfcaton J.A. ouwelse,. Garback, D.H.J. Epema, H.J. Sps (2004), A Measurement Study of the BtTorrent eer-to-eer Fle-Sharng System, Techncal Report DS-2004-003, Delft Unversty of Technology, The Netherlands. aurent Bernalle, Renata Texera (2006), Traffc classfcaton on the fly [J], ACM SIGCOMM Computer Communcaton Revew, Vol36(2), pp23-26. Olver Heckmann, Axel Bock (2002), The edonkey 2000 rotocol, ftp://ftp.kom.e-technk.tu-darmstadt.de/pub/ papers/hb02--paper.pdf. ablo Rodrguez, See-Mong Tan, Chrstos Gkantsds (2006), On the feasblty of commercal, legal 2 content dstrbuton [J], ACM SIGCOMM Computer Communcaton Revew, Vol36(), pp75-78. Savage S., Collns A., Hoffman E. (999), The End-to-End Effects of Internet ath Selecton [C], roceedngs of ACM SIGCOMM, pp289-299. Sarou S, Gummad K, Grbble S D (2002), A measurement study of peer-to-peer fle sharng systems [C], roc. of the Multmeda Computng and Networkng 2002, pp56 70. Stemm M., Katz, R., Seshan S. (2000), A network measurement archtecture for adaptve applcatons [C], INFOCOM 2000, Vol(), pp285-294. Sen S, Wang J (2004), Analyzng peer-to-peer traffc across large networks [C], IEEE/ACM Transactons on Networkng, Vol(2):2, pp29-232. Vnay Aggarwal, Anja Feldmann, Chrstan Schedeler (2007), Can ISS and 2 users cooperate for 2023