Mean Waiting and Turnaround Time for Multiple Web Transaction

Transcription

1 Mea Waiig ad uraroud ie for Muliple Web rasacio Yog-Ji Lee Dep. of Copuer Sciece Woosog Uiversiy aejo, , Korea M. Aiquzzaa School of Copuer Sciece Uiversiy of Oklahoa Nora, OK 7309, USA Keywords: ea waiig ie, ea aroud ie, web rasacio, H, C, -C Absrac Hyper ex rasfer roocol (H) is a rasfer proocol used by he World Wide Web disribued hyperedia syse o rerieve objecs. H is a coecio-orieed proocol which uses C (rasissio Corol roocol) as he raspor layer proocol. H does o ierac very well wih C. I his paper, we sudy he facors ha affec he perforace of H over C ad he rasacio ies of C ad -C. We iroduce wo ew perforace crieria, ea waiig ad uraroud ies, as easures of Qualiy of Service (QoS) of uliple ed users accessig web servers ad develop he aalyical odels providig he heoreical upper boud for such QoS. Resuls fro our proposed odel ach well wih experieal daa. he odel ca be applied o diesioig of ework lik badwidhs ad disribuio of web servers o saisfy a give QoS o ed users as easured by he ea waiig ad uraroud ies.. INRODUCION he World Wide Web uses he Hyper ex rasfer roocol (H) a he applicaio layer o rasfer objecs bewee copuers i he Iere. H uses o he rasissio Corol roocol (C) [2] as he raspor layer proocol. H over C suffers fro several probles such as he coecio seup, slow-sar, ie_wai delay ad sall MSS (Maxiu Sege Size) [3,4,]. A uber of soluios o he above probles arisig due o he use of C as he uderlyig raspor proocol for H have bee repored i he lieraure. For exaple, -C [6,7] reduces he ie_wai delay fro 240 secods o 2 secods ad also decreases he rasacio ie of C (which is a leas 2 R file rasissio ie) o R file rasissio ie by eliiaig he hree-way coecio seup delay. I he case of persise H [8] (icludig H/. [9], H-NG [], ad S-H []), rasacios followig he firs oe do o require ew coecios o be seup, ad herefore, avoids slow-sar o each ew rasacio. here have bee a uber of sudies [-4,, 6] o he perforace probles of H. hey copue he rasacio ie, akig io accou he ework badwidh, widow size ad slow-sar. However, he previous sudies are liied o sigle user cases, where he rasacio ie is a good easure of he ed o ed delay experieced by a user. he oivaio of his paper is o sudy he case of uliple users accessig a server, where he waiig ad uraroud ies deped o he server load. I such a case, he rasacio ie ay o be a good easure of ed o ed user delay. he focus of his paper is o deerie he user level perforace of H i he case of uliple users. he resuls repored i his paper ca be used by ework egieers o diesio a ework i ers of badwidh requiree ad o develop schee disribuig he load aog a uber of web servers i order o iprove he QoS perceived by ed users. he objecive of his sudy is o fid he heoreical upper boud of he acual waiig ad uraroud ies of users i a real eviroe. We achieve our objecives by developig a aalyical odel for such QoS ad validaig he odel agais real world experies o a es bed. I coras o previous work, which oly cosidered he rasacio ie, wihou akig io accou he load o he server or uliple users, he resuls of his paper will allow us o copue he ed-o-ed QoS experieced by a user i he case of uliple users. he auhors are o aware of ay previous paper which have used he ea waiig ad uraroud ies as QoS of ed users. he coribuios of his paper are as follows: (i) Developig a aheaical fraework o copue he ea waiig ad uraroud ie of a ed user whe uliple users siulaeous access he server, (ii) Diesioig of ework lik badwidh ad disribuio of web servers o saisfy ed user QoS he res of he paper is orgaized as follows. I secio 2, we aalyze he rasacio ies for C ad -C respecively ad derive aalyical expressios for he ea SECS '04 7 ISBN:

2 waiig ad uraroud ies. Secio 3 discusses ad copares resuls fro real world experies, followed by cocludig rearks i secio MODELING AND ANALYSIS I his secio, we discuss he ea waiig ad uraroud ies ha are he os ipora QoS facors perceived by ed users. Mea waiig ad uraroud ie are defied as he average ie whe users eed for he sar ad he fiish of web service respecively whe uliple users access he web server siulaeously. ouch e. als [] easured he iefficiecy of a H proocol whe ier-workig o C. hey aalyzed he ieracio ie of H ad copued he coecio ie per rasacio, ad a upper lii o he addiioal slow sar ie ad copared i wih he os suiable rasissio ie wihou slow sar. Also, hey assued ha he processig capabiliy of a hos is oly affeced by he badwidh of a ework; hey igored he processig ie of server ad disk ipu/oupu boleecks. Moreover, hey do cosider he ea waiig ad uraroud ies ha are ipora crieria for ed o ed QoS defied i his paper. If he ea waiig ad uraroud ies desired by users are give, we ca fid he iiu badwidh or server disribuio sraegy o saisfy ed-users QoS. We defie he followig oaios ha will be used i our odel i secio 2 ad eriology ad Noaios R: R represes he roud rip ie bewee he clie ad he server. bw: badwidh represes he rasissio speed of lik (bps: bi per secod) MSS: axiu sege size (packe size) K: he uber of packes i he file which is give by K filesize MSS L: he uber of packes i oe widow, ad is give by L bw R MSS M: axiu useful widow size (lower boud). I is give by he saller of he uber of packes fillig he pipe durig a R a he axiu available widow size ad he uber of packes i he file. I, herefore, represes he axiu uber of packes ha ca be rasied a ay ie, ad is give by i (L, K). S: uber of roud rips durig C's slow-sar (upper boud). C s iiial sedig widow sars a oe MSS. However, i os BSD ipleeaios, i is icreased o wo o receip of ACK for he SYN packe. he widow for daa raspor, herefore, sars a oe ad icreases by power of wo for each roud rip ie. I is herefore give by S log2 M / 2. he slow sar ie is give by R S. Fro Figure, he acual file rasissio ie is he su of he iiu file rasissio ie ( ) ad he addiioal delay (R S) due o slow sar. Noe ha alhough R S has o be icluded as par of he ie for C Daa (reply), i is depiced i he upper par of he figure i order o avoid cofusio wih i o be discussed laer. W: wased ie. I is he su of oe roud rip ie for coecio seup ad a os oe roudrip ie for slow-sar. I is give by W slow sar ie coecio seup ie R S R. : iiu file rasissio ie wihou cosiderig slow-sar. I is obaied by dividig he by he badwidh, i.e. filesize bw. However, because C uses slow-sar i a real eviroe, his echais will lii he iiial widow o oly oe packe uil he firs ACK is received, ad subsequely icrease he widow gradually by power of 2. hus, eve if he ework has eough badwidh, he clie cao use he full badwidh uil he widow grows large eough, i.e. he acual file rasissio ie will be loger due o he ifluece of slow-sar. i : iiu rasacio ie. I is he su of Ror H reques ad he iiu file rasissio ie wihou slow sar ie. I is give by i R. : rasacio ie for C. I is give by i W 2R R S. -C : rasacio ie for -C. I is obaied by excludig he coecio seup ie (R) i ad is expressed as -C R R S. w : oal waiig ie : oal uraroud ie w : ea waiig ie : ea uraroud ie he followig addiioal oaios are used whe all packe rasacio ies are sae. : oal uber of users : he uber of packes requesed per user : rasacio ie per packe (/ or -C /) he followig addiioal oaios are used whe packe rasacio ies are differe. : he uber of differe ypes of liks depedig o heir badwidhs, such as, E, ec. i : he uber of users usig badwidh ype i(,2,) : oal uber of users as give by i i i : rasacio ie per packe usig badwidh ype i SECS '04 72 ISBN:

3 2.2 H over C ad -C H uses oe C coecio for a rasacio ad does o eed ay corol chael i he server or he clie. he ielie for file rasissio i H is show i Figure. Clie Nework Server () Web server processes packes i a roud robi fashio. (2) All users reques he sae file, i.e. he uber of packes per user is sae (3) Server processig ie is igored. For exaple, CU ad disk I/O perforace of he server are o cosidered. W i R S R R file received C SYN C SYN, ACK C ACK H reques C Daa (reply) Figure. ielie for file rasissio usig H over C We have already discussed several probles icurred as a resul of he use of C by H. As show i Figure, C us se up a coecio ad execue he cogesio corol echaiss such as slow-sar before sedig he H reques. I Figure, he os desirable rasissio (opial rasissio) will o iclude he coecio seup process ad cogesio aagee echais. ha is, i is o sed H reques ad o receive a file wih he axiu rasissio speed correspodig o he badwidh (bw). Figure 2 shows he ielie for file rasissio usig H over -C. I Figure 2, he reaso ha he posiio of R S is appeared i he firs par of he figure is sae as he reaso i Figure. Clie R S R -C i Nework C SYN, FIN H reques C SYN, FIN, ACK Server 2.3 Aalysis of ea waiig ad uraroud ies I his secio, we develop aalyical odels for he ea waiig ad uraroud ies for wo cases depedig o wheher he packe rasacios ies are sae or o. he rasacio ie () foud earlier is oal ie for a user o coec o a web server ad dowload a file. Sice we assue per packe roud robi for web service ad he dowload cosiss of several packes, packe rasacio ie, / (assuig sae ie for all rasacios). Also, is give by () / MSS. Now, if here are uliple clies requesig he sae file fro web server, each user's expeced rasacio ie () will be he sae. However, he acual fiishig ie of clie will be affeced by he uber of users who are accessig he server siulaeously. Whe he packe rasacio ies () of users are sae, he ea waiig ad uraroud ies are give by Eqs. (2) ad (4) respecively. O he oher had, whe he rasacio ies per packe are differe, he ea waiig ad uraroud ies are give by Eqs. () ad (6) respecively Case : Sae packe rasacio ies Whe he web servers are coeced o he exeral users hrough oly oe lik, he oal waiig ie, he ea waiig ie, oal uraroud ie ad ea uraroud ie are give by he followig equaios: w ( i) ( )( ) () i ( i) ( )( ) i w (2) file received C Daa (reply) [ ( ) ] (3) 2 Figure 2. ielie for file rasissio usig H over -C We ake he followig assupios for developig he odel for he ea waiig ad uraroud ie: 2 [ ( ) ] [ ] (4) Case 2: Differe packe rasacio ies Whe he web servers are coeced o he exeral users hrough ay liks of differe badwidhs, he ea waiig ad uraroud ie are give by Eqs. () ad (6) SECS '04 73 ISBN:

4 respecively. Firs, we cosider he ea waiig ie. o fid he waiig ie of i h user, we divide he oal ie io wo iervals: he firs ierval represes he ie whe all he packes excep he las packe of each user has bee received; he secod ierval represes he ie whe he las packe of each user has bee received. oal waiig ie of i h user uil he firs ierval is (he uber of packes - ) [(he uber of users for group icludig i h user ) i (oal packe rasacio ie excludig i h group)]. he waiig ie of i h user is he su of rasacio ies of oher users prior o hi. By geeralizig ad addig his all, we obai he followig equaio for he ea waiig ie. Boh 0 ad 0 are zeros i he equaio. i i ( ) i [( i ) i j j]] [ i ( j j ) ( j ) i ] i i, j i i j j w Now, we cosider he ea uraroud ie. If we use he sae procedure as he waiig ie, oal uraroud ie of i h user uil he secod ierval is (he uber of packes - ) [he uber of users ( i ) he su of packe rasacio ie ( i )]. he uraroud ie of ay user i he secod ierval is he su of rasacio ies of oher users prior o hi ad his ow packe rasacio ie. hus, by geeralizig ad addig his all, we obai he followig equaio. Boh 0 ad 0 are zeros i he equaio. i i ( ) i i [ i ( j j ) ji i i j j ] Case 3: Disribued servers If he web servers are disribued over he idividual lies, we firs fid he ea waiig ad uraroud ies by applyig Eqs. (2) ad (4) o each lie idividually. Sice he uber of differe badwidh ypes is assued o be, we add up idividual ea waiig ad uraroud ies ad divide by. Hece, wo ea values are sae as Eq. (7) ad Eq. (8) respecively. k ( k i) k k ( k )( k ) k k w w [ i ] / (7) k k k k k ( k ) [ k ( k ) ] k 2 k I he equaio, k w, k () (6) (8), k, k, ad k represe he ea waiig ie, he ea uraroud ie, he uber of users, he uber of packes, ad he packe rasacio ie of he k h lie respecively. 3. ERFORMANCE EVALUAION I our experie, we coeced a H server (Liux 6. OS / Apache.3.x Web Server) ad clies (SUN ulra- WS / Solaris 2.6 OS) direcly hrough Mbps Ehere. We used files sizes of 3000, 6000 ad 2000 byes. MSS was 460 byes ad R (R) was easured as 0. s. We coposed he socke ierface usig UNIX C laguage i he clie, ad varied he uber of users (,, ad ) wih he sae uber of hreads as he uber of users. able shows resuls obaied by applyig eriology ad oaios i secio 2. o he experieal eviroe of which badwidh Mbps, R 0. s ad MSS 460 byes. he reaso ha i ad -C are he sae i he able is because S was calculaed as zero i he experieal seup. If S is o zero, -C becoes larger ha i by R S as show i secio 2.2. (bye) K L (s) M S W (s) i (s) (s) able. Resuls fro he aalyical odel i he experieal eviroe -C (s) able 2 ad 3 show he copariso of ea waiig ad uraroud ies of he proposed odel ad experie for sae ad differe packe rasacio ie respecively. ad -C of he copuaio colu i able 2 have bee reproduced fro able ; he correspodig values i he experie colu are he average of resuls obaied by execuig he above C progra e ies for he sae size file ad he uber of users. I able 2, he uber of packes () represes he ceilig of value (K) ha divided io MSS. Each packe rasacio ie () is a value o divide oal rasacio ie () io uber of packes (). We obaied w ad by applyig,, o Eq. (2) ad Eq. (4) of secio For he differe packe rasacio ies, we divide he users io groups of fiveuser ( ) ad e-user ( 2 ) for each. I he copuaio colu of able 3, ( i ) ad -C ( i ) are reproduced fro able, ad hose i he experie colu are he iiu aog values obaied by execuig he C progra e ies. w ad were foud by usig Eqs. () ad (6) i secio respecively. he reaso ha he copuaio values are less ha he experieal oes is because he server processig ie i he C progra is icluded i he laer. We ca ifer ha he SECS '04 74 ISBN:

5 differece bewee he copuaioal ad experieal values represes he processig ie. able 2. Copariso of resuls bewee odel ad experie for sae packe rasacio ies (All ies are i illisecods) (byes) (bye) copuaio [Eq. (2) Eq. (4)] C -C () w -C () w (.) (0.99) (.8) (.08) (.20) (.) experie C -C () w -C () w (4.77) (4.7) (3.7) (3.6) (3.96) (3.90) able 3. Copariso of resuls bewee odel ad experie for differe packe rasacio ies (All ies are i illisecods) (bye) (bye) user/ group ( i ) ( i ) w user/ group 3.46 (.).92 (.8).83 (.20) copuaio [Eq. ()~Eq. (6)] C -C C ( i ) ( i ) w 8.0(6.0) 3.0(4.3) 22.0(4.4) 7.0(3.4) 38.0(4.2) 36.0(4.0) -C ( i ) (0.99) (.08) w (.) experie -C -C ( i ) (.8) 2.(4.2) (4.3) 6.(3.3) (4.2) 3.(3.9) w able 4 shows he rasacio ie for several badwidhs. We assued badwidhs as (.44 Mbps), E (2.048 Mbps), 3 ( Mbps), ad AM OC-3 ( Mbps) ad used 2 byes as he MSS of C i WAN. R is he easured value i he dedicaed lie o ouer WAN ad i was 0. s. As for he, 6 KB were used ad hese values ea leghs of files sored i web server. Of course, we ay copue while chagig he above paraeers. K, L,, M, S, W, i,, ad -C copued o each badwidh are sae as able 4. able 4. rasacio ie for several badwidhs badwidh (Mbps) (.44) file size (bye) K L (sec) M S W i (sec) (sec) (sec) -C (sec) M E (2.048) M (44.736) M AM-OC () o see he effec of server disribuio, we cosidered six groups while chagig he uber of users fro o 0. Also, we assued wo s, ad E, wo E s, ad 3, E ad 3, ad ixed, E, 3 for he cobiaio of differe lies. he uber of users ( i ) for each group were assued as half of oal uber of users whe he uber of lies is wo ad a raio of 3:3:4 of hose whe he uber of lies is hree. ie divided oal rasacio ie () io he uber of packes () i able 4 was used as he packe rasacio ie ( i ) of each group. able represes he resul whe he is 6,000 byes ad MSS is 2 byes. able. Mea waiig ad uraroud ie wih he ixed badwidh eviroe Group badwidh i a a' b b' each uber E of lies is oe * 2 2/lie (.) (.36) E (9.80) (.23) 3 E * 2 2/lie (9.66) (.08) (9.3) (9.94) E (9.39) (9.79) E (6.20) (6.62) Now, we defie savigs rae i case of server disribuio as he Eq. (9). a( a') b( b') Savigs Rae 0 (%) (9) a( a') SECS '04 7 ISBN:

6 I Equaio, a, a', b, ad b' represe he ea waiig ie wihou server disribuio, he ea uraroud ie wihou server disribuio, he ea waiig ie wih server disribuio, ad he ea uraroud ie wih server disribuio respecively. a, a' are obaied by applyig Eq. (2) ad Eq. (4) respecively. b, b' are obaied by applyig Eq. (7) ad Eq. (8) i secio By applyig he above-eioed procedure o MB ad 6 KB files, we foud a, a', b, ad b' respecively. he, by Eq. (9), we obaied Figure 3 showig he savigs rae (%). I is kow ha i Figure 3, he savigs rae of he ea waiig ie is superior o ha of he ea uraroud ie i boh 6 KB ad MB files ad becoes up o 0 ~ perce geerally. Figure 3. Effec of server disribuio 4. CONCLUSIONS he explosive growh i dead for web service i rece years has icreased he dead for he Iere. However, he real access speed experieced by ed users is droppig as copared o he early Iere. Alhough H, a represeaive proocol of web, was desiged regardless of he uderlyig ework srucure, C/I is he predoia uderlyig proocol. herefore, H has o ierac wih he C layer. Sice he wo proocols do' ier-work opially, a uber of probles such as several hudreds of s of delay i coecio seup, srog ifluece of slow-sar due o he sall size of web files, ad cosa value of he ie_wai saus i coecio release reduce he ed o ed perforace of users. I his sudy, we firs exaied perforace probles of H over C, ad he copued he rasacio ie o each badwidh while cosiderig widow size ad slow sar. We derived aalyical odels o copue he ea waiig ad uraroud ie for siulaeous access of he web server by several users. his sudy focused o he perforace degradaio of he C rasacio ie as a fucio of he icreasig uber of users. he objecive of sudy is o provide he heoreical upper lii abou he real eviroe whe we assue he ea waiig ad uraroud ie as QoS of ed-user. Our aalyical odel has bee validaed hrough experieal resuls. We foud ha he ea waiig ie ca be reduced by 0% by properly disribuig he load o he web servers. he resuls repored i his paper ca be used o fid aleraives abou how o disribue aou of badwidhs i order o saisfy he ea waiig ad uraroud ies experieced by ed-users. O he oher had, we cofired ha he waiig ad uraroud ies of -C are beer ha ha of C, ad i becoes ore efficie as he or he badwidh becoes saller. Fuure exesio of his work icludes perforace aalysis of oher iproved C proocols ad eviroes wih sochasic rasacio ies. Refereces [] Berers-Lee, J., Fieldig, R. ad Frysyk, H. 99. Hyperex rasfer roocol (H/.0), RFC-94. [2] osel, J. 98. rasissio Corol roocol. RFC-793 / SD-007. [3] Seves, R. 99. C/I Illusraed, Volue 3: C for rasacios, H, NN, ad he UNIX Doai roocols, Addiso-Wesley. [4] Lee, Y he sudy o he age for direcory i e- coerce. Korea Elecroics elecouicaio Research Isiue: 6-8. [] Rescorla. E. ad Schiffa. A he secure Hyperex rasfer roocol [6] Brade. R Exedig C for rasacios- Coceps, RFC-379. [7] Brade. R /C - C Exesios for rasacios: Fucioal Specificaio. RFC-644. [8] Mogul. J. 99. he Case for ersise-coecio H. I roceedigs of he ACM Sigco '9: [9] Fieldig. R. Frysyk. H. ad Lee Hyperex raspor roocol (H/.), RFC-266. [] Sio. E. 99. Nex Geeraio Hyperex raspor roocol, Iere Draf. [] Heidea. J. Obraczka J. ad ouch. J Modelig he erforace of H over Several raspor roocols, IEEE/ACM rasacios o Neworkig,, o. : [2] Heidea. J erforace Ieracios bewee -H ad C Ipleeaios, ACM Copuer Couicaios Review: [3] Jacobso. V Cogesio Avoidace ad Corol, I roceedigs Of he ACM Sigco '88. [4] Jacobso. V. ad Karels. M Cogesio Avoidace ad Corol, ACM CCR, 8, o.4: [] Mogul, J. ad Deerig, S ah MU Discovery, RFC-9, DECWRL, Saford Uiversiy. [6] adaabha, V. ad Mogul, J Iprovig H Laecy, I roceedigs of he Secod Ieraioal WWW Coferece. SECS '04 76 ISBN: