Optimal Video Distribution Using Anycasting Service

Transcription

1 Bond Unverty Informaton Technology paper Bond Bune School Optmal Vdeo Dtrbuton Ung Anycatng Servce Zheng da Wu Bond Unverty, Chr oble Bond Unverty Dawe Huang Follow th and addtonal work at: Recommended Ctaton Zheng da Wu, Chr oble, and Dawe Huang. (999) "Optmal Vdeo Dtrbuton Ung Anycatng Servce".Jun Th Conference Paper brought to you by the Bond Bune School at It ha been accepted for ncluon n Informaton Technology paper by an authorzed admntrator of epublcaton@bond. For more nformaton, pleae contact Bond Unverty' Repotory Coordnator.

2 Optmal Vdeo Dtrbuton Ung Anycatng Servce Z.D. Wu and C. oble School of Informaton Technology Bond Unverty Gold Coat, Queenland 49,Autrala Telephone: ; Fax: E-mal: D. Huang School of Mathematc Queenland Unverty of Technology Garden Pont Campu, GPO Box 434 Brbane Qld 400, Autrala Telephone: ; E-mal: Abtract In the Internet, a group of replcated erver commonly ued n order to mprove the calablty of a networked ervce. An anycatng ervce the new network ervce that reolve an anycat addre to one t member IP addre from the group. The anycatng ervce ue ome crtera to chooe the one "bet" erver a a detnaton. In th paper we wll preent an optmal algorthm for mappng each anycatng query from clent nto the one "bet" vdeo dtrbuton erver of replca, uch a Vdeo-On-Demand erver on the Internet. The algorthm developed at applcaton-layer of the network baed on economc model and queung theory. By ung th algorthm clent can get maxmal atfacton whle ytem reource can be utlzed mot effcently. Keyword: Anycatng; Economc Model; Optmal Load Dtrbuton; VoD Server. Introducton In order to ncreae ervce avalablty and provde effcent load dtrbuton n a network, t common practce to replcate erver on the network today. Example of uch ervce nclude World-Wde-Web "mrror" te, Vdeo-On-Demand, SOCK erver [4], compute-erver and proxy erver n the Internet. Th often referred to a

3 the calablty of a ervce. There have been dfferent approache propoed for mprovng the calablty of a network ervce, uch a cachng, batchng of requet at the erver, multcatng of erver repone, and erver replcaton over the network. In th paper we only deal wth anycatng mechanm that can upport erver replcaton. The noton of anycatng n the Internet wa frt ntroduced by C. Partrdge, T. Mendez, and W. Mllken n [6]. Anycatng alo defned n IP veron 6 (IPv6) [7]. Bae on the defnton of anycatng n [6][7], we may undertand th ervce from everal apect a follow: () an IP anycat addre (one of uncat IP addree) ued to dentfy a group of erver that provde the ame ervce; () an anycat meage the one that hould be delvered to one member of a group of degned recpent or erver and preferably to the "bet" member. () t dfferent from multcat n whch a member mut be ent to every member n the group. There are two notable reearch report n anycatng, whch repreent two dfferent approache to workng on th area. In report [5], the author degned and mplemented a network-layer anycatng ervce for load dtrbuton n the context of the IBM Olympc Web te on the Internet. They nvetgated how the IP anycat ervce can be exploted by hot connected to the Internet wthout gnfcantly mpactng the routng and protocol proceng nfratructure already n place. The author propoed poble enhancement to routng and forwardng to fully explot the potental of anycat ervce. In another report [8], applcaton-layer anycatng communcaton paradgm wa propoed to upport erver replcaton. Specfcally, the author developed and evaluated an mplementaton baed on the ue of anycat reolver to map anycat doman name nto one or more IP addree. The 3W-erver replcaton choen a a target to ue anycatng ervce n both of the work. In th paper, we wll preent an optmal algorthm for mappng each anycat requet (or query) nto one "bet" vdeo dtrbuton erver of the replca, for example, a group of Vdeo-On-Demand erver whch provde the ame vdeo tream ervce on the Internet. Our work developed at applcaton-layer and followed the noton preented n [8]. The algorthm degn theoretcally baed on economc model and queung theory. By ung th mechanm, clent or uer wth

4 Qualty of Servce (QoS) requrement can obtan maxmal atfacton, the reource of erver replcaton can be utled effcently and therefore the whole ytem can acheve the bet performance. Th paper organed a follow. Followng the decrpton of applcaton-layer anycatng n Secton, we ntroduce an optmal algorthm for vdeo dtrbuton ung the anycatng ervce n Secton 3. In Secton 4, we dcu economc model and optmzaton technque, the theoretcal bae on whch the algorthm propoed and then, demontrate ome numercal reult. Fnally, a concluon and further dcuon about th ubject gven. Sytem Decrpton In th ecton we wll follow the noton preented n [8] to brefly decrbe applcatonlayer anycatng. An anycat doman name (AD) ued to dentfy a group of erver that provde the ame ervce. We conder a collecton of VoD erver that may be located n dfferent place on the network and offer vdeo tream ervce. An anycat doman name unquely dentfe a et of IP addree, whch form an anycat group. Such group can be made up entrely of uncat IP addree or entrely of multcat IP addree. In both cae, anycatng ervce provde a mappng from the AD repreentng the group to the "bet" IP addre n the collecton accordng to ome crtera. Conceptually, each of the IP addre n the group repreent the addre of one replcated erver and reachng any one of them acceptable, but n practce one of mportant problem how to realze an anycatng ervce to fnd out the "bet" IP addre or memeber. The bac motvaton of th paper to try to anwer th queton by developng an optmal algorthm for makng uch decon n the context of vdeo dtrbuton ervce. The anycatng requet/repone cycle hown n Fgure a below.

5 Clent Anycat query Reolver Server The bet IP addre Vdeo tream A clent ntate an anycat query wth an anycat doman name and t addtonal requet-pecfcaton, uch a QoS parameter, the reolver procee the query by executng a decon algorthm and reple wth the "bet" IP addre a t anycat repone. The anycatng archtecture lke the extng Doman ame Sytem (DS) nfratructure n the Internet [0]. Logcally, we conder ue related to anycat reolver degn eparately from the DS. In realty, the functon of anycat reolver could be ntegrated wth operaton of the DS[8]. ow, let u dcu what the meanng of the "bet" erver from among multple replcated erver. We may nvetgate th problem from two apect a follow: () from the pont of vew of clent, they wh to obtan ther maxmal atfacton for acceng the ervce, for example, they wh to gan the mnmal call blockng probablty under ther QoS requrement; () from the pont of vew of a whole ytem, t hould acheve t bet performance whenever a repone IP addre mappng reolved. In other word, the ytem reource hould be utled mot effcently under the "bet" ervce choen. The quanttatve expreon for the uer atfacton and the ytem performance wll be gven n Secton 4. 3 Algorthm In th ecton, an optmal algorthm for mappng each anycat requet nto the "bet" vdeo dtrbuton erver of the replca gven. It theoretcal bae and detal of

6 optmzaton procedure wll be dcued n the next ecton. The algorthm cont of two module, whch are mplemented n the reolver and each erver of a group repectvely, and llutrated n Fgure a below: A query from clent Reolver Server ID {Spec: λ, d, q } S { VFD } Server λ VFD A A* λ B C q Reply to the query For the reolver, a table or data tructure contructed to contan ytem meaure and charactertc referenced by each erver ID, n the table. λ repreent the average arrval rate of the AD query at erver wth type of ervce,. A denote an uncat IP addre agned to erver. repreent the value for decon (VFD) durng the mappng proce whch generated from erver for each of ervce type. q denote a QoS parameter of ervce type whch can be lnked to λ (not hown n the dagram). Whenever an AD query arrve at the reolver, t wll be n a queue. The reolver wll execute the mappng proce for the quere watng n the queue under the FCFS dcplne. For each query of type, the reolver wll fnd out the maxmum of among all of erver, max,, and the correpondng erver, *. Conequently, the IP addre, * A, can be choen a the "bet" repone to t clent. Another functon of the module n the reolver to update the value of entre n the table. At each tme of updatng, the module wll determne a new value of λ and end

7 them to each erver. Once each erver receve th value, the erver trggered and mmedately compute t new value of the VFD,, a t repone to the reolver. The reolver wll update the VFD n the table mmedately after each of them arrve n. The econd module wll be mplemented n each erver. The erver alo need a data tructure to keep nformaton for generatng the VFD value once t trggered by the reolver. Bede λ, q and addtonal meaure are needed and they are: for each type of ervce at erver and defned a the ame charactertc a above, two B - the current free buffer pace avalable C - the current avalable bandwdth of network output lnk for erver wth type. Each erver wll run an optmzaton procedure whch take B, C, λ and q a nput value and produce the VFD,, a t output value, whch wll be ent to the reolver mmedately after t generated each tme. The detal of how to determne the VFD wthn each erver and how and when to update λ wthn the reolver wll be gven n the next ecton. 4 Model and Optmzaton In order to fnd out the "bet" erver for each anycat query from among multple replca, an optmal decon algorthm propoed n the lat ecton. The algorthm theoretcally baed on analytc technque n economc and queung theory. By ung th algorthm, the ytem reource are utlzed mot effcently, meanwhle the uer (or clent) can gan ther maxmal atfacton. The detal of mathematcal model ued n the algorthm now preented a the followng. Suppoe that a vdeo dtrbuton ytem on the network cont of S erver, each of them may be dentfed a, =,,..., S. Each vdeo erver provde concurrent vdeo tream or connecton wth dfferent type of ervce, =,,..., I. Each type of vdeo tream aocated wth t dplay peed at d frame per econd, whch requeted by t end-uer or clent. We defne the value for decon (VFD),, a the number of

8 potental connecton whch can be addtonally upported by erver. It note that doe not repreent the number of connecton the erver have et up. Aumng that anycat query arrval of ervce type a Poon proce wth average rate λ whch wll be further dvded by the algorthm nto arrval ub-rate λ for each erver. Thu, S λ = = denote λ. The rule of th dvon wll be dcued at the end of th ecton. We b, meaured n frame, a buffer ze to be needed for etablhng connecton of new uer of ervce type wth erver and c n frame per econd a bandwdth requrement for the ame reaon. Thu, I b = B and I c C, where B and repreent the total of free buffer pace and the total of avalable bandwdth of network output lnk wth erver repectvely. Suppoe that the vdeo tream are ent over the network from the erver at the ame rate a they dplay on the end-uer devce. Therefore, c = d. C In practce, uer prefer to mnmze ther call blockng probablty. Thu, uer atfacton or preference may be meaured by the call blockng probablty, q. A a reult, a utlty functon repreentng the uer atfacton expreed a It reaonable to aume that U = q = q. q ndependent of the dentfer of a replcated erver. The call blockng due to the hortage of ytem reource, uch a buffer and bandwdth wth a erver. For a erver the clent requet, whch wa refued by the erver, often referred to a "lot" due to lmted reource avalable n the erver. We may employ M/M//B queung model [] to evaluate the lo rate, δ, a follow: δ = b, p k k = 0 where

9 λ / c p, k = b ( λ / c ) 0 + ( λ / c ) k 0 k b otherwe Gven the preference of all uer wth dfferent type, we can aggregate ther utlty functon nto a ocal welfare functon, whch may repreent the performance of a erver a whole []. There are two method to defne the ocal welfare functon for th optmzaton problem: () we ue a ocal welfare functon to drectly reflect the call blockng probablty of uer requeted for the ervce; () we employ a ocal welfare functon to repreent the total number of potental connecton, I =, whch can be addtonally upported by erver. Logcally, thee two method are equvalent n term of the meanng of choong the "bet" erver a dcued n Secton. In th paper we wll ue the econd method and thu, we may defne the ocal welfare functon for each erver a follow: W ( U, U,, U I ) I = = Then, we can poe the welfare maxmzaton problem: Max W U, U,, U ( I ) Such that U = q I b = I c = B C δ

10 where 0 q <. The frt contrant mple that the ue preference of each type mut be atfed under a certan lmtaton. Specfcally, q repreent the call lo rate thoe uer of type may be tolerated. In optmzaton problem n economc, a very neat technque called "the method of Lagrangan multpler" generally ued. For the cae of two type of uer, =,, we eek to maxmze the ocal welfare functon above, by ung the Lagrangan theorem we can obtan the followng equaton: c Φ b log ρ Φ b = c log ρ B = b + b C = c + c Where λ / c ρ =, λ / c ρ = and Φ ρ ( q ) = ρ ( q ) Through thee equaton we can acheve a et of Pareto allocaton of buffer and bandwdth for uer of each type, { b } c, [][][3]. Baed on th optmal reource allocaton we can ealy etmate the maxmum number of potental connecton, max,, for each type of uer, that can be addtonally upported by erver ung, = c / d. max A mulaton carred out baed on the algorthm for optmal reource allocaton wth erver, a dcued above. In the mulaton, we expre the utlty functon for each type a U = q and =,. It mean that we aume two type of uer compete for reource (buffer pace and network output lnk capacty) n the numercal analy. Suppoe that the buffer pace and the network output lnk capacty of erver are frame and B = 80 C = 00 frame/ repectvely. The requet arrval rate from clent to

11 erver are λ and λ whle the frame output rate or frame dplay rate are c = d and c = d. Certanly, the dplay rate depend on ther ervce type. The typcal example of vdeo data type nclude 5 frame/ anmaton, 30 frame/ TSC vdeo, and 60 frame/ HDTV vdeo. Baed on thee pecfcaton we can now determne optmal buffer allocaton, b, under call lo contrant, U = q δ. We let q = 3 0 and q 4 = 0 or 0. Once we get reource allocaton atfyng the call lo contrant we can etmate the maxmum number of tream or connecton the erver can upport addtonally. We aume d = 5 frame/ and d = 30 frame/ whch repreent two typcal vdeo dplay rate, and reman all aumpton above. The numercal reult are frt hown n Table and Table where they have dfferent value of q. In thee two table, B mn (certanly, B mn B ) repreent the mnmum buffer pace requred o a to acheve the maxmum number of the addtonal connecton denoted a for type. Thu, + the total number of connecton that can be added to the erver. Accordng to the numercal reult, we remark that () f the vdeo demand (or requet) arrval rate λ ncreae, the buffer ze to be allocated alo ncreae; () the call lo contrant mpact on the buffer ze to be allocated but they do not affect the maxmum number of addtonal connecton very much. Let u examne Table 3 where we reman buffer pace B = 80 frame and ncreae the network output lnk capacty up to C = 00 frame/. Comparng wth Table and Table, th change lead the maxmum number of addtonal connecton ncreaed gnfcantly meanwhle the allocated buffer ze become maller. Table : Aumng B = 80, C =00, d =5 and d =30.

12 Table : The call lo contrant are et to dfferent value from Table. Table 3. Aumng B = 80, C =00, d =5 and d =60. In order to update the metrc n the mappng table, we mut have an algorthm that the reolver wll ue to determne when the arrval rate of anycat query for each type of ervce, λ etmated baed on the traffc meaurement from clent. A uggeted n [8], we alo borrow and adopt the lnk tate update algorthm ued n the ARPAET [9]. The update algorthm parameterzed by a meaurement nterval L, a maxmum threhold T and a reducton factor R. The algorthm mantan a current threhold M, ntalzed to T. The reolver meaure t tate over each nterval L. If the tate change from the prevou meaurement by at leat M, the tate puhed and M reet to T. If the tate doe not change by at leat M, M reduced by R. The algorthm wll produce at leat every T/R tme unt and at mot every L unt. Baed on th updatng algorthm,

13 λ can be frt etmated at each tme of updatng. Then, we can compute λ for each erver by ung the expreon a follow: λ = = I = ( λ ) It mean that f a erver can upport more number of potental connecton, the reolver wll return t addre to more number of clent. Concluon We have preented an optmal algorthm for mappng each anycat requet (or query) nto one "bet" vdeo dtrbuton erver of the replca n the context of a group of Vdeo-On- Demand erver, whch provde the ame vdeo tream ervce on the Internet. Our work developed at applcaton-layer. The algorthm degn theoretcally baed on economc model and queung theory. By ung th mechanm, clent or uer wth Qualty of Servce (QoS) requrement can obtan maxmal atfacton, the reource of erver replcaton can be utlzed effcently and therefore the whole ytem can acheve the bet performance. The numercal reult provde ueful nght n the performance behavor of a vdeo erver. It noted that f the number of ervce type large, the cot of computaton for optmal reource allocaton wll be hgh. However, th wll not caue too many overhead n the ytem nce the erver uually equpped wth a hghperformance computer. Acknowledgment The work decrbed n th paper wa funded by the Autralan Reearch Councl Grant. Reference [] Hal R. Varan, "Intermedate Mcroeconomc - A Modern Approach", (Second Edton), 990.

14 [] S. Charle Malrce and Owen R. Phllp, "Economc Analy: Theory and Applcaton", (Ffth Edton), 986. [3] D. F. Ferguon, C. kolau and Y. Yemn, "An Economy for Flow Control n Computer etwork", Proc. of the IEEE IFOCOM, 990. [4] M. Leech, M. Gran, Y. Lee, R. Kur, D. Kobla, and L. Jone, "SOCKS protocol veron", RFC 98, 5 Aprl 996 [5] Erol Baturk, Robert Engel, Robert Haa, Vnod Per and Debanjan Saha, "Ung etvork Layer Anycat for Load Dtrbuton n the Internet", IBM Reearch Report, RC 0938 (07997), IBM Reearch Dvon, T. J. Waton Reearch Center, ew York. [6] C. Partrdge, T. Mendez, and W. Mllken, "Hot anycatng ervce", RFC 546, ovember 993 [7] R. Hnden and S. Deerng, "IP veron 6 addreng archtecture", RFC 884, December 995. [8] S. Bhattacharjee, M. Ammar, E. Zegura, V. Shah, and Z. Fe, "Applcaton Level Anycatng", Tech. Rep. 96-5, College of Computng, Georga Inttute of Technology. [9] E. C. Roen, "The updatng protocol of ARPAET' new routng algorthm", Computer etwork, o. 4, pp.-9, 980. [0] P. Mockapetr, "Doman name - concept and faclte ", RFC 034, ovember 987. [] L. Klenrock, "Queucng Sytem", Vol., Vol., Wley, 976. [] Z.D. Wu, "Dynamc Seon Control for Multmeda Dtrbuted Sytem", Proc. of IEEE ICC'97, pp