Massive Data Delivery in Unstructured Peer-to-Peer Networks with Network Coding

Transcription

1 Yu, Shu and L, hongwen 7, Massve data delvey n unstuctued pee-to-pee netwoks wth netwok codng, n IEEE/ACIS Intenatonal Confeence on Compute and Infomaton Scence (ICIS 7), IEEE Compute Socety, Los Alamtos, Calf., pp IEEE. Pesonal use of ths mateal s pemtted. Howeve, pemsson to epnt/epublsh ths mateal fo advetsng o pomotonal puposes o fo ceatng new collectve woks fo esale o edstbuton to seves o lsts, o to euse any copyghted component of ths wok n othe woks must be obtaned fom the IEEE.

2 Massve Data Delvey n Unstuctued Pee-to-Pee Netwoks wth Netwok Codng Shu Yu, hongwen L School of Engneeng and Infomaton Technology, Deakn Unvesty, VIC 35, Austala Emal: syu@deakn.edu.au Infomaton Scence and Technology College, amen Unvesty, amen 365, P. R. Chna Emal: lzw@mu.edu.cn Abstact Wth moe and moe multmeda applcatons on the Intenet, such as IPTV, bandwdth becomes a vtal bottleneck fo the boomng of lage scale Intenet based multmeda applcatons. Netwok codng s ecently poposed to take advantage to use netwok bandwdth effcently. In ths pape, we focus on massve multmeda data, e.g. IPTV pogams, tanspotaton n pee-to-pee netwoks wth netwok codng. By though study of netwokng codng, we ponted out that the peequstes of bandwdth savng of netwok codng ae: ) one nfomaton souce wth a numbe of concuent eceves, o ) nfomaton peces cached at ntemedate nodes. We futhe poof that netwok codng can not gan bandwdth savng at mmedate connectons to a eceve end; As a esult, we popose a novel model fo IPTV data tanspotaton n unstuctued pee-to-pee netwoks wth netwok codng. Ou pelmnay smulatons show that the poposed achtectue woks vey well.. Intoducton Multmeda applcatons ae boomng n the Intenet, and thee s a ctcal challenge fo utlty of cuent avalable Intenet bandwdth. Fo eample, TV ove Intenet Potocol (IPTV n shot) [8] s an emegng and nevtable tend of the TV ndusty n the nea futue, whch s to delve dgtal televson pogams va the wdely connected wold wde web facltes. IPTV allows people to customze the own TV pogams and ft the pogams nto the pesonal tmetables. Howeve, avalablty of netwok bandwdth becomes a bottleneck fo the boomng of TV ove Intenet. The cuent ntemedate nodes, e.g. outes, n the Intenet povde vey lmted effots to mpove the bandwdth effcency The ecently ndusty effots on content dstbuton netwoks (CDN) [4], [6] s an applcable soluton fo the challenge. In CDN, the content of dffeent applcatons s dstbuted to dffeent local domans befoe the possble equests ae ssued. It s obvous that CDN educes the pessue of bandwdth utlzaton, howeve, thee ae two ctcal poblems n CDN: pedcton of the possble equests and challenge of stoage space fo the CDN seves. One applcable soluton fo the challenges s applyng pee-to-pee netwoks [3]. It s a beakthough of the tadtonal clent seve model, and completely dstbutes the esouces and sevces, theefoe, t has a vey stong scalablty and hgh pefomance. Thee ae a numbe of vey successful klle applcatons ae based on pp systems, Fo eample, the Skype, a pp based hgh qualty global voce ove IP sevce. Cuently, pp netwok s the best choce fo IPTV, and thee ae some applcatons, such as, Based on CDN, we have a numbe of eplcas of one fle n the Intenet, theefoe, paallel downloadng [], [3], [4] s a natue way to conduct data tanspotaton. [] eseached the mult-seve contnuous meda documents delvey ove the Intenet. It poposed a mechansm to delvey dffeent pats of a desed meda fle to a clent fom dffeent nstalments, theefoe, the delay of clent watng tme wll be mnmsed. The advantage of paallel downloadng s that t can damatcally educe the watng tme of a specfc use, although thee ae aguments [], [5] about the pefomance f all the nodes apply paallel downloadng stategy concuently n a system. Howeve, CDN tansfe the challenge on netwok 6th IEEE/ACIS Intenatonal Confeence on Compute and Infomaton Scence (ICIS 7) /7 $5. 7

3 bandwdth to the challenge on ntemedate nodes stoage capabltes. Netwok codng [] s poposed fo usng the netwok bandwdth effcently. The obsevaton s that the cuent ntemedate netwok nodes, e.g. outes, only eceve packets and then fowad them to the down steam wthout any pocessng on the message tself. The pncple s f outes compess messages that they eceved fom the up steams and then fowad them to the net nodes, then the utlty of bandwdth wll be damatcally mpoved. Pevous papes [], [6] showed that netwok codng does gan bandwdth savng n theoy. Futhemoe, eseache n Mcosoft [7] mplemented a pototype, Avalanche, wth netwok codng fo lage scale content dstbuton systems, the epements showed that applyng netwok codng ncus lttle ovehead n tems of CPU pocessng and I/O actvty, howeve, t esults n smooth and fast downloads. Netwok codng can also mpove the elablty fo IPTV content delvey as well, because thee ae edundant eplcas of nfomaton peces dstbuted at ntemedate nodes. We gve a vey smple eample fo the benefts of netwok codng below. Souces S (,, 3) Intemedate nodes Receves Receve 3 3 Receve Fgue. An eample fo netwok codng applcaton In Fgue, the IPTV content s dssected nto thee segments evenly,,, 3, the codng functon s OR, and we use to denote the length of message, theefoe, = =, =,,3 =,,3 and k =,,3 k Shown as n Fgue, the segments ae coded and dstbuted n the ntemedate nodes, a eceve can download the encoded segments, and econstuct the ognal message,, 3 based on the encoded segments. We suppose, evey edge n Fgue s equvalent, and the bandwdth cost fo each edge fo one message s. Wthout codng, the cost of one eceve s 6 (3+3), and fo two eceves, howeve, wth codng, the cost of two eceves s (4+3+3) <. Moe ove, t s moe elable fo content delvey wth codng than that wthout codng, e.g. eceve can download the message fom,, o 3,, o 3, 3,. 3 Namely, the system s moe elable wth netwok codng The est of ths pape s ogansed as follows: secton touches the elated wok and analyse netwok codng mechansm. A novel model fo massve multmeda data tanspotaton n pp netwoks s poposed n secton 3 based on the analyss n the pevous secton. The smulatons ae dscussed n secton 4. Fnally, secton 5 summases the pape, and pesents the futue wok.. Related Wok and Analyss Netwok codng fst poposed n [], t s a new mechansm n ode to mpove the thoughput utlzaton of a gven netwok. The pncple behnd netwok codng s allow the ntemedate nodes, e.g. outes, to encode packets. As we know, the cuent outes only eceve packets and then fowad them wthout any opeatons on the message n the packets. An ovevew of netwok codng and a dscusson of the potentals ae gven n pape [6]. Lnea netwok codng has been eploed a lot [5], [], []. Fst of all, we eamne the mechansm fo netwok codng. In a dected netwok G ( u, v), u s the set of vete, and v epesents the lnks n the gaph. Thee s a souce node s, and d ( d ) eceves n set T, s T, T s the set of end nodes. The nfomaton to be delveed fom the souce node s ndependent dscete andom segments,,...,, The output segments at a eceve node β ae denoted as ( β, ). Edge l caes the segment Y (l). Fo multcast on a netwok wth lnk delays, the eceves have to possess stoage capablty, although memoyless opeaton suffces at all othe nodes. The coespondng lnea codng equatons ae as follows [9]. Y t+ ( ) = a,, t + fl, Yt ( l) { : geneated at v} { l: head ( l) = v} μ μ t+ ( β, ) = bβ, l t u ( β, ) + bβ, l Yt u ( l) u u u= { l: head( l) = β} u= () () 6th IEEE/ACIS Intenatonal Confeence on Compute and Infomaton Scence (ICIS 7) /7 $5. 7

4 Whee head ( l) = v means dected edge l ponts to vete v; μ epesents the memoy equed; a,, fl,, bβ, l and b u β, l u ae coeffcents of a specfc codng netwok. An llustaton of the codng at an ntemedate node s shown n Fgue. Y() μ V Y(3) Y() = a,3 + f,3y ( ) + f, 3Y () = Fgue : Illustaton of lnea codng at an ntemedate node In Fgue, ( =,..., μ, μ ) ae the nfomaton that have been stoed at node v so fa. a,3, f,3 and f,3 ae coeffcents of the local codng. Fom the system pont of vew, the whole codng netwok can be vewed as a tansfe mat ( A, F, B). Ths mat bdges the nput M β vecto of souce nfomaton [,,..., ] and the coespondng output vecto,,..., ] at one eceve node β : [ β, β, β, [,,..., ] M β ( A, F, B) = [ β,, β,,..., β, ] (3) Whee A elated wth the outgong lnks of the souce s, B elated wth the ncomng lnks to the eceve node β, and F elated wth the sub-gaph between the souce s and the eceve β, n each ntemedate node, the codng of the ntemedate node s gven by equaton () and (). The essental poblem hee s how can we decde the tansfe mat M β ( A, F, B), and then we can econstuct the ognal nfomaton,,..., ] fom,,..., ], and the [ [ β, β, β, codng on the nfomaton futhe educes netwok bandwdth consumpton and mpoves the netwok thoughput and pefomance. Pevous papes have poofed that thee est bandwdth savng usng netwok coded. Howeve, the peequstes ae: One nfomaton souce, at least two concuent eceves fo nstant nfomaton delvey An loose condton fo beneft fom netwok codng n tems of bandwdth savng s the ntemedate nodes hold message peces as long as possble fo non-concuent eceves We use B e ( ) to donate the bandwdth cost of tanspotng nfomaton on edge e. Theefoe, the followng holds Be( ) = B ( ) = B (... ),, k, and e e (4) k k Lemma. Fo any gven eceve, at the mmedate ncomng lnks to the eceve, the least bandwdth consumpton s Be ( ), =,,.., n ode to econstuct the ognal message =,,..., ]. [ Poof We suppose that fo a gven eceve, thee ae m ncomng paths, e, e, e m, then, B m e, e..., e Be ( ), m ( ) = s the encoded message fom,,..., Step. When h=, then = [, ]. In ode to obtan the ognal message, we have to eceve at least elements fom set {,, ). Fo each of the encoded message, the bandwdth consumpton s C = B ( ), =,. Theefoe e B C e e ( ).,, Lemma holds Step. Suppose lemma holds when =n-, then B,,..., ) ( n ) C e e e (,,..., m n Step 3, when h=n, then B e, e e =,..., m, m, (,,..., n, n ) B e, e..., e (,,..., n ) + B e, e..., ( n ) C + C = nc Lemma holds e m ( ) In anothe wod, at the last hop to a eceve, thee s no bandwdth savng fo any pece of message. Fom Lemma, we can conclude that netwok codng gans no bandwdth savng at the last hops to the end eceve. Howeve, t s obvously, the pefomance of fle downloadng s mpoved aound h tmes. Lemma. A message = [,,..., ] s tanspoted and encoded by netwoks, and,,..., ) s the full lst of the possble encoded ( n schema fo a eceve β. In ode to econstuct the ognal message, eceve β only needs any subset of,,..., ) wth dffeent elements. ( n n 6th IEEE/ACIS Intenatonal Confeence on Compute and Infomaton Scence (ICIS 7) /7 $5. 7

5 Poof Wthout loss genealty, = c + c c,,..., = Once we have dffeent packets, theefoe, we have = c + c c = c + c c... = c + c c IPTV Souce Coepees Pees Souce Laye Stoage Laye Receve Laye c c... c And we know ank of mat c c c s, c c... c theefoe, we can obtan vecto = [,,..., ]. As a esult, lemma holds. 3. A Model fo IPTV Data Tanspotaton Thee ae two categoes of pp netwoks: stuctued and unstuctued pp. In the stuctued pp netwoks, thee s a logcal stuctue among pees, theefoe, the pefomance of the stuctued pp netwoks s bette than that of unstuctued pp netwoks. IPTV s a eal tme system, and t possesses a stct equement n tems of delay, hence, the stuctued achtectue s chosen n ou poposal. In ode to educe the netwok taffc and meet the ctea of IPTV pefomance, we oganse the pees nto clustes n tems of geogaphy. A cluste head s selected fo each cluste, whch acts as a tempoay stoage space fo local pees. We name the cluste head as coepee n the contet of ths pape. The selecton of a coepee depends on two factos: lage stoage capablty and bandwdth avalablty to the othe pees. We manage the IPTV pp netwoks n thee layes: the eceve laye; the stoage laye; and the souce laye. All the pees, coepees and the souces ae physcally connected. The pees ae dstbuted n the Intenet. The coepees establsh a logcal laye, the coepee laye o stoage laye. In a wod, all the nodes n the IPTV pp netwok ae connected togethe physcally, and they ae ogansed n a tee achtectue logcally, shown as below. Fgue 3. the thee laye pp netwok achtectue fo IPTV 3. Content dstbuton Once an IPTV pogam s avalable, the souce node evenly dssects the fle nto segments,,,...,. Once thee s a equest fo the fle, the segment,,..., wll be tanspoted to coepees va the outgong lnks of souce node S n tun. Fo a segment, t may encounte any segment,whch s a combnaton of a subset of,,...,, on the way to a coepee. The segments wll be encoded accodng to fomula () and (). Encoded message segment wll be hold n stoage space once t aves a coepee. wll stay n the coepee as long as t can. Because of the lmtaton of stoage capacty, wll be eplaced accodng to specfc ules, such as fst n fst out. 3. Infomaton Reteval Once an IPTV eceve ssues a equest fo a specfc TV pogam, an nfomaton eteval pocedue stats. We use a cawl agent fo message segments eteval. Accodng to lemma, we only need a subset of (,,..., n ) to econstuct the ognal message. the agent goes though the coepees one afte anothe to seach fo the elated message segments o the combnatons. If one sutable element s founded at a coepee, then stats to download the element, at the same tme, the agent moves to net coepee. We set an up bounday, N, fo the eteval at the coepee laye, f the agent can not obtan dffeent elements 6th IEEE/ACIS Intenatonal Confeence on Compute and Infomaton Scence (ICIS 7) /7 $5. 7

6 afte N coepees, then t wll edect the quey to the souce node S fo the emanng elements. The algothm s lsted as follow = {,,... } = { any conbnaton of, =,,..., } L = P = {null} when ( N) //N s the up bounday of eteval steps n coepee laye {// seach any message segment of at coepee P = coepee( )[ q( f ( P = null o P L) { + +, contnue} ; { beak}; } else )] { downloadng(p); L = L + P ; f ( length ( L) = ) //a clent does not have dffeent element dung the N ound // eteval at coepee laye then t downloads the est of the //message segments fom the souce node S If ( length( L) p ) { P = L downloadng(p) } esult s shown as Fgue 4. We found that paallel path downloadng acheves a much bette pefomance than that of sngle path downloadng. When thee ae 6 paths, fo dffeent fle length, the vaaton of tanspotaton delay s vey lmted. It s a geat featue fo data tanspotaton of IPTV fles. T anspo t at on Del ay Fle Length Fgue 4. Tanspotaton delay vs fle length usng paallel downloadng pat h pat hs 4 pat hs 8 pat hs 6 paths Because of the lmtaton of stoage capacty, the encoded message segments can not be hold at one coepee fo a long peod, theefoe, some message segments may mss fom the coepees, we name ths as segment mssng. Fgue 5 shows the tanspotaton delay vesus mssng ate. The smulaton ndcates that mssng ate has slght mpact on the vaaton of tanspotaton delay. T anspo t at on Del ay Fle Length= Fle Length= Fle Length=4 Fle Length=8 4. Smulatons Pelmnay smulatons have been conducted fo the poposed model fo multmeda data tanspotaton n pee-to-pee netwoks wth netwok codng. In the smulatons, tanspotaton delay s used as a metc fo the dffeent cases. We fst eamne the tanspotaton delay fo fou fles, whose fle lengths ae,, 4 and 8 espectvely. Fo each fle, we ethe use one path to tansfe the fle o, 4, 8 o 6 paths to do paallel downloadng. The M ssng Rate Fgue 5. Tanspotaton delay vs mssng ate fo dffeent length of fles Moe smulatons about the poposed model wll be conducted n the nea futue. 6th IEEE/ACIS Intenatonal Confeence on Compute and Infomaton Scence (ICIS 7) /7 $5. 7

7 5. Summay and Futue Wok In ths pape, we pont out eplosvely that ganng netwok bandwdth savng usng netwok codng depends on ethe one souce vesus multple eceves, o stoe message segments at ntemedate nodes. We poved that at the last connects to a eceve, thee s no bandwdth gan wth netwok codng, futhe moe, n ode to econstuct the ognal message,,...,, we poved that a eceve needs only eceve dffeent elements of the andom combnaton of,,...,. Based on the analyss, we poposed a model fo IPTV data tanspotaton n pee-to-pee netwok wth netwok codng. Pelmnay smulatons ndcated that the poposed model woks vey well n tems of massve multmeda data tanspotaton. The cases fo multple dffeent nfomaton souces and lage numbe of eceves scenao wll be eploed as the futue wok. Moeove, compehensve smulatons wll be conducted n the nea futue. Refeences [] R. Ahlswede, N. Ca, S.R. L, and R.W. Yeung, Netwok Infomaton Flow, IEEE Tansactons on Infomaton Theoy,. [] Geassmos Balas, Bhaadwa Veeavall, Optmzed Dstbuted Delvey of Contnuous-Meda Documents ove Unelable Communcaton Lnks, IEEE Tansactons on Paallel and Dstbuted Systems, Vol. 6, No., Octobe, 5. [3] J. Byes, M. Luby, and M. Mtzenmache, Accessng multple mo stes n paallel: Usng tonado codes to speed up downloads, n Poc.IEEE INFOCOM, vol., New Yok, NY, Ma. 999, pp [8] Ramesh Jan. "I Want My IPTV," IEEE MultMeda, vol., no. 3, pp. 96, 95, July-Septembe, 5. [9] R. Koette and M. Medad, An Algebac Appoach to Netwok Codng, IEEE/ACM Tansactons on Netwokng, (5), Octobe 3. [] Smon G. M. Koo, Cathene Rosenbeg, and Dongyan u, Analyss of Paallel Downloadng fo Lage Fle Dstbuton, Poceedngs of the Nnth IEEE Wokshop on Futue Tends of Dstbuted Computng Systems (FTDCS 3), 3. [] S.-Y. R. L, R. W. Yeung, and N. Ca. "Lnea netwok codng". IEEE Tansactons on Infomaton Theoy, Febuay, 3. [] S.-Y. L, N. Ca, R. Yeung, "On Theoy of Lnea Netwok Codng", ISIT 5 [3] Dean S. Mlocc, Vana Kalogeak, etc. Pee-to-Pee Computng, A Hewlett-Packad ntenal techncal epot,. [4] Pablo Rodguez and Enst W. Besack, Dynamc Paallel Access to Replcated Content n the Intenet, IEEE/ACM Tansactons on Netwokng, Vol., No. 4, August. [5]Janto Song, Chaofeng Sha and Hong hu, Nash Equlba n Paallel Downloadng wth Multple Clents, Poceedngs of the 4 th Intenatonal Confeence on Dstbuted Computng Systems (ICDCS 4), 4. [6] D. Vllela and D. Rubensten, Pefomance Analyss of Seve Shang Collectves fo Content Dstbuton, IEEE Tansactons on Paallel and Dstbuted Systems, VOL. 6, NO., Decembe 5. [4] O. Ecetn and L. Tassulas, Maket-Based Resouce Allocaton fo Content Delvey n the Intenet, IEEE Tansactons on Computes, VOL. 5, NO., Decembe 3. [5] C. Fagoul and E. Solann, "Requed Alphabet Sze fo Lnea Netwok Codng", IEEE Intenatonal Symposum on Infomaton Theoy (ISIT 4), June 7th-July nd, Chcago. [6] C. Fagoul, J. L. Boudec and J. Wdme, Netwok Codng: An Instant Pme, ACM SIGCOMM Compute Communcaton Revew, Volume 36, Numbe, Januay 6. [7] C. Gkantsds and P. R. Rodguez, Netwok Codng fo Lage Scale Content Dstbuton, IEEE INFOCOM 5. 6th IEEE/ACIS Intenatonal Confeence on Compute and Infomaton Scence (ICIS 7) /7 $5. 7