Thortical and Applid Informatic ISSN 1896 5334 Vol 20 (2008), no 1 pp 15 27 Mthod and algorithm of brokr-bad HTTP rqut global ditribution LESZEK BORZEMSKI a,anna ZATWARNICKA b,krzysztof ZATWARNICKI b a Intitut of Information Scinc and Enginring Wrocław Univrity of Tchnology Wybrzż St Wypiańkigo 27, Wrocław, Poland lzkborzmki@pwrwrocpl b Intitut of Automatic Control and Computr Opol Univrity of Tchnology ul Mikołajczyka 5, Opol, Poland { azatwarnicka,kzatwarnicki}@poopolpl Rcivd 10 July 2008, Rvid 24 July 2008, Accptd 8 Augut 2008 Abtract: In th papr w prnt nw brokr-bad HTTP (Hyprtxt Tranfr Protocol) rqut global ditribution mthod and algorithm calld GARDIB (Global Rqut Ditribution with Brokr) Th mthod i bad on a fuzzy-nural dciion-making chm allowing ditributing HTTP rqut bad on th rqut rpon tim Th dciion at global lvl i mad by th brokr which i an intrmdiat rvr ud by th Wb clint to gt Wb pag Each HTTP rqut i rdirctd by a brokr which i locatd in th nighbor proximity of th ur Svral brokr may rv to th am t of Wb it with th fully rplicatd contnt Brokr ha ability to timat rqut rpon tim uing own mad ntwork brokr-to-wb it latncy maurmnt, a wll a load condition of vry local Wb it, collctd from local Wb witch that locally ditribut rqut to th targt local Wb rvr In th papr w aum that local Wb witch ar quippd with th FNRD (Fuzzy Nural Rqut Ditribution) algorithm dvlopd by th author FRND dciion-making chm alo i bad on fuzzy-nural modling with rqut compltion tim a th prformanc mtric GARDiB i valuatd via imulation xprimnt and compard with othr candidat for global ditribution, namly RR (Round Robin) andwrr(wightd Round Robin) polici Th wighting in th lattr ca can b bad on th local load tat (WRR L) or ntwork throughput (WRR T) Simulation xprimnt ar mad by man of CSIM19 oftwar in C++ Th rult how that GARDiB can b ffctivly ud for global HTTP ditribution xhibiting good prformanc in th wid rang of load Kyword: HTTP rqut ditribution, WWW rvr, imulation Thi work wa upportd by th Polih Minitry of Scinc and Higr Education undr Grant No N516 032 31/3359 (2006 2009)
16 1 Introduction Modrn Wb it ar bad on Wb Srvr clutr tchnology to provid rliabl and powrful rvic Wb rvr clutr ar quippd in Wb witch lying in th front of th Wbit to ditribut locally incoming HTTP rqut to individual Wb rvr to complt th rqut rvicing Du to clutr tchnology th total load can b ditributd ovr Wb rvr in th Wbit rulting in bttr ovrall prformanc Th rqut ditribution may b improvd by global ditribution organizd for a group of Wbit localizd uually in diffrnt gographical localization in Intrnt Thi approach i wll known in CDN (Contnt Ditribution Ntwork) infratructur Such olution nd to utiliz om global ditribution mchanim to dcid which local Wbit i to b rvd th rqut Thi papr prnt a mthod and fuzzy-nural algorithm of brokr-bad HTTP rqut global ditribution which optimiz rqut rpon tim Our approach i rlatd to ur prcivd prformanc which i valuatd by th rqut rpon rqut tim, not any prformanc indx rlatd to Wb it opration uch a load balancing Fuzzy-nural ytm hav bn ud in many intllignt-bad application Fuzzy controllr dal wll within uncrtainty, noi and non-linarity condition Uing nural nt it i poibl to add lf-larning fatur which can b a bai to dvlop adaptiv ytm Uing a fuzzy and nural hybrid mthod w can mak u advantag of both approach Our prior rarch [5] introducd Fuzzy-Nural Rqut Ditribution (FNRD) algorithm for local HTTP rqut dipatching at local lvl that i for Wbit quippd with a clutr of Wb rvr Thi i th application layr ditribution a it rcogni objct fatur rqutd in pcific HTTP rqut In wa hown in practical tup a wll a through xtniv imulation xprimnt that FNRD ovrcom th ur prcivd prformanc of uch ditinguihd ditribution algorithm lik wll known CAP (Contnt Awar Policy) [11], LARD (Locality Awar Rqut Ditribution) [2, 15], and RR (Round Robin) [8] algorithm which ar oftn dicud in th litratur a good on Th papr prnt th architctur and organization of rqut ditribution at global and local lvl, ditribution algorithm, a wll a imulation rult Our ytm i calld Global Adaptiv Rqut Ditribution with Brokr (GARDiB) and mploy fuzzy-nural dciion-making approach to minimiz rqut rpon tim at both lvl, global and local W aum vry local Wb it ha th am contnt and may rv vry rqut
17 2 Architctur of th ditribution ytm Architctur of th HTTP rqut global ditribution ytm i hown in Figur 1 Th ytm conit of a numbr of local Wb rvr clutr (LS Local Srvic) mploying om local ditribution algorithm, clint and brokr In thi papr w aum that all local Wb witch ditribut HTTP rqut according to FNRD algorithm [4] with th fuzzy-nural dciion-making chm Thi aumption i important a that thr would b poibl to mploy any othr local rqut dipatching algorithm - uch ca i to b tudid in othr our work lwhr Clint INTERNET Brokr #3 Local Srvic 3 Fig 1 Architctur of ditribution ytm Th clint nd th rqut to thir nighboring brokr that rdirct th rqut to th Wbit for which th timatd rqut rpon tim i timatd a th lat on Th nighbor brokr i chon on th bai th hortt ditanc maurd by th ntwork dlay In particular, it could b localizd to rv om givn ur domain Th brokr hould b localizd clo to th clint that upport a clo a poibl to mak poibl ntwork dlay btwn clint and brokr ngligibl Thr ar a many a rquird brokr to mak advantag of poibl load balancing Th rqut rpon tim tak into account tim to pnd to tranfr th rqut and th rqut rply, a wll a th tim of rqut compltion by LS Mor prcily, it i calculatd for th brokr, not for particular clint, a th tim pnd by th brokr from tarting tranfr of th firt byt of th rqut to th nd of tranfr of th lat byt of th rplay Th architctur of th brokr i hown in Figur 2 It conit of th following main componnt: Claification Modul, Local Srvic Modl Modul, Dciion Making Modul, Maurmnt Modul and Excution Modul
18 S a i a i 1 a i ~1 t i Claification modul k i SL 1 x i 1 qi qi Maurmnt modul S q i Local Srvic modul SL S S tˆi Local Srvic modul SL tˆ i Dciion making modul w i Local Srvic modul SL 1 ˆi1 t Excution modul x i SL SL S Brokr Local Srvic Fig 2 Architctur of brokr All incoming rqut ar claifid to on of pr-dfind cla k, k=1,2,ě, K by th Claification Modul Thi claification i Wb objct iz bad in ca of th tatic rqut and rqut contnt in ca of th dynamic on Th ditinguihing btwn tatic and dynamic rqut will b dicud in th imulation ction of thi papr Each brokr upport all of th S local rvic LS 1, LS 2,,LS,,LS S Each Local Srvic Modl (LSM) Modul work to th bnfit of a givn ingl LS Th ultimat aim of LSM i to timat th rqut rpon tim for th rqut i ca of rvicing it by th LS upportd by givn LSM Thi tim for th i-th rqut x i and timatd for th -th LS i dnotd by ˆti Thi tim includ ntwork dlay and Wb it rvic tim LSM of th -th LS mploy th fuzzy modl to timat th rpon tim for i-th rqut taking into account th following information: k i th cla of th rqut, ai currnt local load which i givn by th numbr of TCP connction opn and activly rvicd by -th LS, and qi tranfr tim (calld ntwork load) of a prob packt nd btwn givn brokr and -th LS Th local load a 1 i, a2 i,,ai,,as i ar collctd by all brokr from Wb witch oprating at vry local Wb it of th brokr Th fuzzy modl will b prntd latr, togthr with th prntation of local dciion-making approach Th Dciion Making Modul lct th w i th LS, w i {t,,s } for which th timatd rpon tim ha th lat valu Thn th Excution Modul rdirct that i-th rqut to w i -th LS Th rqut rplay tim ti i maurd whn th rqut i compltd by particular LS Th brokr Maurmnt Modul maur rqut rpon tim for all rqut and u thi knowldg to updat brokr dciion-making Figur 3 how th architctur of LSM It conit of thr bai modul, namly Etimation Mchanim, Load Stat Databa, andadaptation Mchanim Th indx of
19 ~ t i Adaptation mchanim k i U k i U k ( i+1) Load Stat Databa U i a i q i U k i Etimation mchanim tˆi Fig 3 Architctur of Local Srvic modul -th LS wa omittd for implification purpo Similar architctur i contructd in ca of global dciion-making prformd by th brokr In th ytm, th timation and adaptation mchanim ar ralizd a a fuzzynural ntwork, which i contructd on th bai of fuzzy-nural modl with Fuzzification Layr, Rul Layr and Dfuzzification Layr Th input data to th fuzzy-nural ntwork ar: ntwork load q,loadr i, cla k i and U ki data from Load Stat Databa U i Load Stat Databa contain actual information about procing in particular LS Bad on thi information th timation modul timat rpon tim for rqut x i Aftr mauring th rpon tim ˆt i, th adaptation mchanim updat th load tat databa U i by changing information about k-th cla from th tat ki to th tat U k (i + 1) Appropriat fuzzy t wr dfind for ach fuzzy-nural ntwork input and alo for th output Th valu of all fuzzy-nural ntwork paramtr ar tord in U i = [U 1i,,U ki,, U Ki ] T Each vctor contain information about rqut blongd to k-th cla at th momnt of rqut x i arrival to th brokr Th paramtr ar ud a th nuron wight and ar tund during th adaptation proc Whn th rqut rvic i compltd, th rqut tim t i can b maurd, which i nxt ud to tun th paramtr of th fuzzy t Thi tuning which procd vry tim aftr compltion of rqut rvic i ralizd uing th back propagation larning mthod 3 Simulation Th imulation wa prformd by man of CSIM19 [12] imulation oftwar CSIM19 i a a library of routin for u by C or C++ programmr; th routin can b ud to implmnt proc-orintd and dicrt-vnt modl Thi oftwar ha bn wll provn in many Wb-rlatd imulation by diffrnt author Our modl ar bad on quuing thory [13] Th Simulator conit of th following main modul: rqut gnrator modul, brokr modul, Intrnt modul and local rvic modul
20 Figur 4 how th baic chm of th imulator Rqut gnrator i to imulat clint rquting for ithr tatic or dynamic WWW rourc W aum in our imulation prntd in thi papr that 20% of all rqut ar tatic that i to gt objct which ar tord a th fil Particular dynamic rqut formulat diffrnt databa load that ar SQL-lik databa tranaction Th rqut from clint can b nt to Local Srvic ithr dirctly or via brokr Dirct acc crat th background load of LS Th paramtr of traffic gnrator hav bn dfind bad on th litratur [6, 7, 10] Each clint i an individual CSIM proc Th clint targtd to th WWW it of th total iz of 200 MB Rqut nt by th clint arriv particular brokr that rdirct it to chon LS faturing th functionality dcribd i th prviou ction Th rol of th brokr modul ar thr fold Firt it rdirct rqut to LS according to th global ditribution algorithm implmntd in it Scond, it monitor th tat of th imulatd ntwork, LS and dploy ndd tatitic Third it rturn th rplay from LS and rturn it to th clint which iud particular rqut W aumd that th dlay clint-to-brokr i ngligibl in both dirction Th brokr modul imulat GARDiB algorithm, a wll a RR (Round Robin) and WRR (Wightd Round Robin) algorithm for comparion nd WRR ha two mod rlatd to prformanc mtric: WRR TandWRR L, whr th formr i ntwork tranfr bad whra th lattr i local load bad WRR aociat to LS a dynamically valuatd wight that i proportional to chon prformanc mtric valu LS W b Srvr Databa rvr LS #1 CPU Clint INTERNET Wb witch DISK LS #S Brokr Fig 4 Simulator tructur In th wightd round-robin WRR L ditribution mod ach LS ha aignd a wight that indicat th procing capacity LS with highr wight fatur with l load and obtain mor rqut firt than tho with l wight Load i dynamically changing and dpnd of th numbr of rqut rvicd by givn LS In th wightd round-robin WRR T ditribution mod ach LS ha aignd wight that indicat ntwork tranfr capacity to that LS from givn brokr LS with highr
21 wight obtain mor rqut firt than tho with l wight Ntwork tranfr capacity i ttd with a prob nt rgularly btwn givn brokr and all LS Th round-robin ditribution RR algorithm aign rqut to all LS in round-robin mannr It trat all LS a qual rgardl of th local load or rpon tim ach LS i xprincing Intrnt modul i aimd at brokr-to-ls and vic vra data tranfr bad on Intrnt modl [14] W xprincd wgt utility [17] and pcap library [16] with additional ral-lif maurmnt on Intrnt to gathr ntwork dlay ditribution to chon Wb it including Autralia, th Nthrland and USA Th xprimnt providd data for imulation of Intrnt path prformanc charactritic In th xprimnt w maurd th TCP connction tablihmnt tim in clint-to-www rvr communication Mor pcifically, it wa th tim intrval btwn nding th TCP SYN packt by th clint and rciving th TCP ACK packt nt by th rvr in rpon to prior SYN packt W ud thi tim intrval a th RTT (Round-Trip Tim) ntwork dlay Th xprimnt wr alo providd ntwork throughput data Local rvic modul i to modl th rvic tim charactritic for th rqut W modlld local functionality faturd Wb witch, WWW rvr and databa rvr A it wa mntiond arlir w imulatd FNRD (Fuzzy-Nural Adaptiv Rqut) local rqut ditribution algorithm [6] A in [2] w modlld procor, dic and cach mmory quu of WWW rvr That modl ha bn ffctivly ud in prviou rarch [5, 9] Howvr du to match modrn computr tchnologi w modrnizd it paramtr by lowring tn tim ach it original tim paramtr Th original tudy in [2] dalt with Pntium II 300 MHz PC, FrBSD 226 oprating ytm and Apach 133 Wb rvr Thu, w ud th following tim paramtr to charactriz th rvicing of HTTP rqut Th CPU tim ndd to analyz HTTP rqut wa aumd 145 µ whra th tim to tranfr 512 B data block wa aumd 4 µ Dic bhavior i charactrizd by th arch and latncy tim 28 m, data tranfr tim 41 µ pr ach 4 KB data block, additional arch and latncy tim for fil gratr than 44 KB 14 m pr ach 44 KB data block Th cach mmory mployd th Lat Rcntly Ud (LRU) algorithm Local databa rvr ar modld a ingl quu [9] Th compltion tim of dynamic rqut wr modld uing th hyprxponntial ditribution [1] 4 Simulation rult Thi ction how and dicu th imulation rult To how GARDiB fatur w propod to load th brokr and local rvic by diffrnt workload In thi papr w xprincd with idntical local rvic quippd with th am numbr of WWW and th am numbr of databa rvr Th global ytm ha bn quippd with a
22 ingl brokr W introduc th following notation of th imulation xprimnt tup dcribing condition of a particular imulation run Th workload how it i aignd to particular LS i dcribd by th ordrd lit of thr trm: th LS country, th numbr of WWW rvr in local Wb rvr clutr, and th har in th total workload dirctly rvicd by that LS A hr w hav only on brokr thr i no nd to dcrib it localization W only pcify th har iz of th total load it rv Thn, for xampl, th notation NL,3,25%; USA,3,25%; AU,3,25%; Brokr,25% dfin th imulation xprimnt tup with thr LS, localizd in th Nthrland (NL), USA and Autralia (AU), ach Wb it i quippd with 3 WWW rvr, and har 25% of th total workload ach In addition, th brokr ditribut 25% of th total load Th workload dirctly rvicd by LS i th background load, a thi traffic w cannot ditribut by brokr Th clint ar localizd in Poland All imulation xprimnt wr prformd to how how particular tup work vru th incraing numbr of clint pr cond Th following prformanc mtric wr tudid in prntd imulation xprimnt: 1 Avrag HTTP rpon tim; 2 95-prcntil of pag latncy tim for all Wb pag acc via a brokr For xampl, 95-prcntil of pag latncy tim i th pag latncy tim limit that th Wb it guarant with 095 probability Th pag latncy tim i th total tim to download th pag klton and all mbddd objct a) b) 1 ] [ 0,8 tim 0,6 n o p 0,4 r t' u 0,2 q r 0 ] 25 [ tim 20 n o p 15 r g a 10 f p o til n 5 rc -p 0 5 9 RR WRR_L WRR_T GARDiB 280 310 340 370 400 430 460 490 520 550 580 610 640 670 clin t / RR WRR_L WRR_T GARDiB 280 310 340 370 400 430 460 490 520 550 580 610 640 670 clint /
23 c) d) 1 ] [ 0, 8 tim 0, 6 n o p r 0, 4 t' u q 0, 2 r ] [ tim rc -p 5 9 0 15 n o 10 p r g a p f o 5 til n 0 RR WRR_L WRR_T GARDiB 280 310 340 370 400 430 460 490 520 550 580 610 640 670 clint / RR WRR_L WRR_T GARDiB 280 310 340 370 400 430 460 490 520 550 clint / Fig 5 Rult for variou imulation xprimnt tup: (a) avrag rpon tim v numbr of clint for < NL,3,25%; USA,3,25%; AU,3,25%; Brokr,25%> tup; (b) 95-prcntil of pag latncy v numbr of clint for <NL,3,25%; AU,3,25%; USA,3,25%; Brokr,25%> tup; (c) avrag rpon tim v numbr of clint for < NL,3,10%; USA,3,40%; AU,3,20%; Brokr,30%> tup; (d) 95-prcntil of pag latncy v numbr of clint for <NL,3,10%; USA,3,40%; AU,3,20%; Brokr,30%> tup Figur 5 how th imulation rult for diffrnt imulation xprimnt tup and workload cnario, namly for two diffrnt brokr load, namly 25% and 30% of th total load, and two diffrnt workload LS har cnario Th dottd lin in Figur 5b and 5d i th thrhold of th 8 c patinc tim howing how long common ur may want to wait for a Wb pag [11] It i obrvd that th ur uually rign to acc a Wb pag if a longr than 8 cond waiting tim xit Th rult prnt that GARDiB ovrcom othr algorithm Th brokr prform wll for both vn (Figur 5a and 5b) and unbalancd (Figur 5c and 5d) load ditribution Good pag latncy rult i achivd for WRR T algorithm whn th LS localizd in th nart ntwork ditanc to Poland, which i in th Nthrland, i lightly loadd (10%) by th background traffic Howvr it prform worn for thi algorithm whn th whol rvic i highly loadd WRR L prform quit wll for all rvic work-
24 load cnario, but whn LS in th Nthrland i highly loadd, and th workload for th whol rvic i quit high, thn Wb pag latncy xcd th thrhold of 8 cond RR algorithm prform wll for non-havy workload, imilarly to WRR T; howvr whn th load incra th rpon tim i unaccptabl 5 Concluion W prntd GARDiB mthod and algorithm for global ditribution of HTTP rqut in contnt dlivry ytm GARDiB utiliz a fuzzy-nural dciion-making modl and how adaptiv and lf-larning capabiliti It ditribut HTTP rqut according to th rqut rpon tim what i pcially vry important for contnt dlivry ytm GARDiB i brokr-bad and HTTP rqut ar rvicd with th hlp of intrmdiary rvr calld brokr Th imulation nvironmnt ha bn programmd uing CSIM19 and C++ oftwar Diffrnt imulation tup and workload cnario hav bn ttd Th imulation rult howd good GARDiB prformanc in comparion to othr ditribution polici Rfrnc [1] MF Arlitt, R Fridrich, T Jin: Workload charactrization of a Wb proxy in a cabl modm nvironmnt, ACM Prformanc Evaluation Rviw, 27(2): pp 25-36, Aug 1999 [2] M Aron, P Druchl, W Zwanpol: Efficint upport for P-HTTP in clutrbad Wb rvr Proc of Unix Annual Tchnical Confrnc, pp 185-198, 1999 [3] N Bhatti, A Bouch, A Kuchinky: Intgrating ur-prcivd quality into Wb rvr dign Proc of th 9th Intrnational World-Wid Wb Confrnc, pp 1-16, May 2000 [4] L Borzmki, K Zatwarnicki: Uing adaptiv fuzzy-nural control to minimiz rpon tim in clutr-bad Wb ytm Procding of 3rd Atlantic Wb Intllignc Confrnc, Poland, Łódź Lctur Not in Computr Scinc, Springr, Brlin, pp 63-68, 2005 [5] L Borzmki, K Zatwarnicki: Prformanc valuation of fuzzy-nural HTTP rqut ditribution for Wb clutr Proc of 8th Intrnational Confrnc Artificial Intllignc and Soft Computing - ICAISC 2006 Lctur Not in Computr Scinc, Springr, pp 192-201, 2006
25 [6] L Borzmki, A Zatwarnicka, K Zatwarnicki: Global adaptiv rqut ditribution with brokr Proc of 11th Intrnational Confrnc on Knowldg-Bad and Intllignt Information & Enginring Sytm, Lctur Not in Computr Scinc 4693, Springr, pp 271-278, 2007 [7] v Cardllini, E Caalicchio, M Colajanni, M Mamblli: Wb witch upport for diffrntiatd rvic ACM Prf Eval Rv, Vol 29, No 2 (2001), pp 14-19 [8] V Cardllini, E Caalicchio, PS Yu: Th tat of th art in locally ditributd Wb-rvr ytm ACM Computing Survy, vol34, no 2, pp 263-311, 2002 [9] E Caalicchio: Clutr-bad Wb Sytm: Paradigm and Dipatching Algorithm PhD Thi [@:] http://wwwcuniroma2it/publication/phdthicaap Univrit? di Roma Tor Vrgata, 2002 [10] E Caalicchio, M Colajanni: A clint-awar dipatching algorithm for Wb clutr providing multipl rvic Proc of 10th World Wid Wb Confrnc, Hong Kong, pp 535-544, May 2001 [11] E Caalicchio, S Tucci: Static and dynamic chduling algorithm for calabl Wb rvr farm Proc of 9th Euromicro Workhop on PDP, Italy, pp 369-376, 2001 [12] CSIM [@:] http://wwwmquitcom [13] T Czachórki: Modl koljkow w ocni fktywności ici i ytmów komputrowych Wydawnictwo Pracowni Komputrowj Jacka Skalmirkigo, Gliwic, 1999 [14] DA Mnac, VAF Almida: Capacity planning for Wb prformanc Mtric, Modl, and Mthod Prntic Hall, Nw York, USA, 2002 [15] VS Pai, M Aron, G Banga, M Svndn, P Druchl, W Zwanpol, E Nahum: Locality-awar rqut ditribution in clutr-bad ntwork rvr Proc of 8th ACM Confrnc On Architctural Support for Programming Languag and Oprating Sytm, pp 205-216, 1998 [16] Pcap [@:] http://wwwtcpdumporg/pcaphtm [17] Wgt [@:] http://wwwgnuorg/oftwar/wgt/
26 Mtoda i algorytmy globalnj dytrybucji ż adań HTTP Strzczni W artykul przntowana jt nowa mtoda i algorytm globalnj dytrybucji żądań HTTP w ici WWW nazwana GARDIB (ang Global Rqut Ditribution with Brokr) charaktryzującgo ię pcyficznym rozmyto-nuronalnym chmatm dcyzyjnym oraz czam odpowidzi jako krytrium jakości dytrybucji Mtoda i algorytm mogą znalźć zatoowani w rozprozonych ytmach dotarczania trści wbowych zwanych iciami CDN (ang Contnt Dlivry Ntwork) Sytm opira woj działani na utalonj w mtodzi wpółpracy klintów, rwiów wbowych oraz pośrdników rwrów pośrdniczących, zwanych brokrami Klinci ą końcowymi użytkownikami ytmu, i w omawianym ytmi aźródłm żądań HTTP Srwiy wbow zwan Srwiami Lokalnymi (ang Local Srvic LS)ą rozprozonymi i autonomiczni działającymi klatrami ztawów rwrowych o kontrukcji: rwr WWW rwr bazodanowy, wypoażonymi w rozdzilacz iciowy (przłącznik wbowy) Srwry pośrdnicząc ą głównymi lmntami ytmu Ich roląjtpośrdniczni w komunikacji klint rwr WWW, co jt ralizowan w natępujący poób Użytkownik kiruj woj żądania HTTP za pośrdnictwm brokra zlokalizowango w najbliżzym ąidztwi Zakłada ię, ż opóźnini czaow wytępując w tj komunikacji jt pomijani mał Brokr korzytając z algorytmu dytrybucji globalnj przkazuj żądani do ralizacji do jdngo zośrodków lokalnych, gdzi rozdzilacz iciowy dokonuj otatczngo wyboru rwra doclowgo Odpowidź od tgo rwra wraca do klinta poprzz lokalny rozdzilacz oraz brokr W pracy proponuj ię, aby dytrybucja globalna była ralizowana z wykorzytanim algorytmu GARDiB działającgo wg rozmyto-nuronalngo chmatu dcyzyjngo, który każd z przychodzących żądań kiruj do tgo lokalngo rwiu wbowgo, dla którgo zotał wytymowany najkrótzy cza odpowidzi Jt to cza obrwowany z punktu widznia brokra Na cza odpowidzi kłada ię cza nizbędny do przłania żądania od brokra do rwiu lokalngo, cza obługi lokalnj żądania oraz cza tranfru odpowidzi od ośrodka lokalngo do brokra Możmy przyjąć, ż jt to równiż cza odpowidzi z punktu widznia klinta końcowgo, z względu na założoną lokalizację brokrów, których moż być odpowidnio dużo zlokalizowanych bliko klintów, np w ich domnach Spcyfiką tj mtody jt zatoowani po raz pirwzy w litraturz przdmiotu podjścia rozmyto-nuronalngo do globalnj dytrybucji żądań HTTP oraz wykorzytani czau odpowidzi jako krytrium wydajności, co jt równiż unikatowym indkm w problmi dytrybucji żądań HTTP, gdzi zazwyczaj toowan ą indky charaktryzując poziom obciążnia rwrów, a clm dytrybucji jt wtdy np równoważni obciążńrwrów W ytmi globalnj dytrybucji żądań GARDiB, w ninijzym artykul proponuj ię, aby w ośrodkach lokalnych działał autorki lokalny algorytm dytrybucji FNRD (ang Fuzzy- Nural Rqut Ditribution), który funkcjonuj równiż wg koncpcji trownika rozmytonuronalngo z krytrium czau wykonania żądania, i który wilokrotni wykazywał ięwoimi dobrymi właściwościami w przprowadzonych badaniach W zczgólności, wyzdł on
27 obronnąręką w porównaniach z znanymi algorytmami lokalnj dytrybucji żądań, takimi jak algorytm karuzlowy RR (ang Round-Robin), LARD (ang Locality-Awar Rqut Ditribution) oraz CAP (ang Contnt-Awar Policy) Dla potrzb przprowadznia badań porównawczych w brokrz zaymulowano karuzlowy algorytm globalnj dytrybucji żądań: RR oraz jgo uogólnionąwrję z wagami, czyli algorytm WRR (ang Wightd Round-Robin) W tym otatnim przypadku badan a dwa warianty algorytmu karuzlowgo ważongo, a mianowici wariant WRR L algorytm karuzlowy ważony z względu na obciążni Srwiów Lokalnych oraz wariant WRR T algorytm karuzlowy ważony z względu na tranfr na trai pomiędzy LS a rwrm pośrdniczącym Do badań wykorzytan zotało środowiko ymulacyjn, tworzon w języku C++ z wykorzytanim bibliotk CSIM19, łużących do programowania zadań ymulacji zdarzń dykrtnych Artykuł pokazuj, ż mtoda oraz algorytm GARDiB mogąbyć z powodznim wykorzytywandoglobalnjdytrybucjiżądańwrozwiązaniachz rwramipośrdniczącymi, orazż ytm globalnj dytrybucji trści wykorzytującyzaproponowanąmtodę oraz algorytm dotarcza woim użytkownikom żądan trści w zadawalającym czai krótzym niż w przypadku pozotałych klaycznych algorytmów dytrybucji