Q-SAC: Toward QoS Opimized Service Auomaic Composiion * Hanhua Chen, Hai Jin, Xiaoming Ning, Zhipeng Lü Cluser and Grid Compuing Lab Huazhong Universiy of Science and Technology, Wuhan, 4374, China Email: hjin@hus.edu.cn Absrac The emerging service grids bring ogeher various disribued services o a marke for cliens o reques and enable he inegraion of services across disribued, heerogeneous, and dynamic virual organizaions. In he experience of consrucing and using he ChinaGrid, we mee wo challenges, opimizing he QoS of he grid resources and minimizing complexiy for applicaion users and developers. In his paper, we presen Q-SAC, a QoS opimized service auomaic composiion model o address hese problems. Two main feaures of Q-SAC are (1 auomaic grid service composiion, and (2 global level mulidimensional QoS opimizaion for he composiion plan. We design he algorihms for generaing and opimizing he composie services. The simulaion resuls show ha our model and soluion are pracical and efficien. 1. Inroducion The emerging grid echnologies have been widely adoped in science and echnical compuing. I suppors he sharing and coordinaed use of diverse resources in dynamic, disribued virual organizaions, ha is, he creaion, from geographically disribued componens operaed by disinc organizaions wih differen policies, of virual compuing sysems ha are sufficienly inegraed o deliver he desired Qualiy of Services (QoS [1]. Recenly, he grid is evolving o he Open Grid Services Archiecure (OGSA [2], which brings ogeher various disribued applicaion-level services o a marke for cliens o reques and enable he inegraion of services across disribued, heerogeneous, dynamic virual organizaions. In our early experience in developing ChinaGrid Supporing Plaform (CGSP [3], a middleware plaform ha addresses building a service-oriened grid based on Web Service Resource Framework (WSRF [4], we found wo major challenges. One is o make he saisfacory use of he grid resources o guaranee QoS. The oher is o minimize complexiy for grid users and developers; ha is, grid-based applicaions should be user-friendly and minimal raining is necessary. To address hese problems, we propose Q-SAC, a QoS-Opimized Service Auomaic Composiion model in his paper. Two main feaures of Q-SAC are (1 auomaic grid service composiion, and (2 global level mulidimensional QoS opimizaion for he composiion plan. The res of his paper is organized as follows. Secion 2 inroduces he semanic based service virualizaion mechanism, which is he main assumpion of Q-SAC. We describe he auomaic service composiion mehod in secion 3. Secion 4 esablishes he mahemaical model for opimizing QoS of he composie plan and proposes he algorihms for soluion. Secion 5 shows he simulaion resuls of Q- SAC. Secion 6 reviews some relaed works. Secion 7 concludes he paper and describes our fuure work. 2. Semanic based Service Virualizaion 2.1 Service Virualizaion Service-oriened view simplifies service virualizaion hrough encapsulaion of diverse implemenaions behind a common service inerface. [5] inroduces a service grid prooype using Legion o explore he design of applicaion and service based replica selecion policies in a wide-area nework. However he semanic of he service canno be * This paper is suppored by Naional Science Foundaion under gran 627376 and 94121, ChinaGrid projec from Minisry of Educaion, and he Naional 973 Key Basic Research Program under gran No.23CB3173.
specified and he replicas of he same service are developed and deployed by he same provider. Our service virualizaion model provides an exensible se of Virual Services (VS o users (see Figure1. Figure 1. Service virualizaion Each VS is defined as uniform service semanics in a specific domain. We call an implemenaion of VS a Physical Service (PS. Thus, each VS is implemened as a vecor of redundan PSs wih differen QoS. From he funcional aspec, each VS may include a vecor of funcions (operaions. Thus we use he following expression o describe a VS: PS O O L O 1 11 12 1v PS O O L O 2 21 22 2v VS = ( F, F F 1 2 v = = (1 L M M M M M PS O O O u L u1 u2 uv where F k denoes he k h funcion of he h virual service and PS k denoes he k h physical service of he h virual service. O ij denoes he j h operaion of h virual service provided by i h service provider. We use Q ij o describe he QoS informaion of O ij: Q ij = Qualiy(O ij (2 and hus he QoS informaion of h VS can be described as he following marix: Q vs =( Q ij; 1 i u,1 j v (3 2.2 Service semanic The semanic of grid services is crucial o enabling auomaic composiion. To help undersand semanic feaures of grid services, we use he concep of onology. Onology is a shared concepualizaion based on he semanic proximiy of erms in a specific domain and is expeced o play a cenral role in he semanic web service composiion [6]. In Figure 2, we describe he proposed onology using a direced graph. Here, nodes represen he conceps of onology. Unfilled nodes refer o WSDL conceps. Gray nodes refer o exended feaure inroduced o augmen WSDL descripions wih semanic capabiliies. In our QoS-aware auomaic service composiion model we pay more aenion o he funcionaliy semanic and QoS semanic of grid service operaions. The node Funcion Rule accuraely describes he funcion of he service in a specific domain. Each operaion has one Funcion Rule [7], which expresses ha he service is capable of producing paricular oupus of he given cerain inpus. The funcion rules are regisered in a rule reposiory when grid services are published. We believe ha as he number of grid services ha offer overlapping or similar funcionaliy increases, QoS will be one of he subsanial aspecs for differeniaing beween similar service providers. The QoS merics used o define QoS capabiliies have o be onologically defined because hey can be very easily misinerpreed [8]. Q-SAC gives accurae definiion of five kinds of applicaion level QoS merics, including price, execuion ime, repuaion, availabiliy, and successful rae. Name Binding Name Inpu Oupu Name 1: (,1 1:n Parameer Type 1:n 1:n 1:n 1: (,1 Uni Funcion Rule Operaion Price Service Purpose Execuion Time Qualiy Repuaion Domain Secuuriy Availabiliy Provider Successful rae Figure 2. Onology descripion of grid services 3. Service Composiion When a developer wans o creae a new composie service, he specifies he inpus and anicipan oupus of he composie service and submis i o Q-SAC. Q- SAC uses he rule engine o deermine wheher he service composiion can be successful. In his secion, we presen he algorihm for generaing he composiion service. The algorihm is rying o deduce he oupu from inpu wih he rule reposiory and i forms he core of he rule engine.
Figure 3. Selecing rules in he rule reposiory We define he inpu se of he user as X, and he oupu of he user as Y, he rule se in he rule reposiory as R. R u denoes he rules ha have been seleced in he process of deducing he oupu. R u is recorded for generaing he execuion pah. The algorihm is described as below: (1 ( X = X, R = Φ, A= Φ ; u (2 ( i ( j rj ( ( ( ( j j j u ( ( j FOR ALL Y X i { IF Y Z r R R z Z z X ( { A AUZj; R R { rj} R R U u u { rj} ; ; ( i IF ( Y ( X U A GOTO ( 5; } ( i+ 1 ( i X X U A; GOTO( 3 ; } (3 ( i 1 ( i IF X = X + i 1 X = X ; GOTO 4 ; ( ( ELSE GOTO ( 2 ; { ( } (4 IF ( Y X + REPORT " No AVAILABLE EXCUTION PATH "; ELSE GOTO ( 5 ; (5 GENERATE PATH WITH R u ; Figure 4. Generaed composiion pah We give he following example for he algorihm. In his scenario, a plaza plans o give birhday gifs o he regisered cusomers for he anniversary celebraion. The ype of gifs depends on hobbies and ranks of he cusomers. Now he manager wans o make sure how much budge will be planned for sending he gifs. For each cusomer, he cusomer ID is inpued o Q-SAC and he oupu of mail service ype wih he price is expeced. Figure 3 shows he rule reposiory and he seleced rules by he generaion algorihm. In Figure 4 we ge a composiion pah. The algorihm selecs necessary rules from he rule reposiory for generaing he composiion pah. The implemenaion should be based on an accessorial rulebased exper sysem. We specify a composiion service in he form of a direced acyclic graph on he assumpion ha cyclic srucure can be ransformed ino acyclic srucure. 4.QoS-opimized Grid Service Composiion In he composiion pah, for each aomic operaion, here may be differen candidae providers wih differen QoS. In his secion, we firs presen he qualiy merics in he conex of elemenary physical service operaion and hen esablish he QoS opimizing model for composie services. 4.1 QoS of grid services We consider five generic qualiy merics for each elemenary operaion: execuion duraion, repuaion, successful execuion rae, availabiliy, and price. The qualiy of a given operaion Q ij can be defined as a vecor: Q ij=(q ij.duraion, Q ij.repuaion, Q ij.suc_rae, Q ij.availabiliy, Q ij.price (4 (1Execuion duraion Given an operaion O ij, he execuion duraion measures he expeced delay beween he momen when a reques is sen and he momen when he resuls are received. In our previous work [9], he execuion duraion is made up of hree pars including compuing ime T com (O ij, middleware ime T mid (O ij, and
ransmission ime T ne (O ij. The execuion duraion is quanified as below: Q ij.duraion = T com (O ij+ T mid (O ij+ T ne (O ij (5 T com (O ij refers o he processing ime of he operaion. If he bandwidh and laency are fixed, T ne (O ij can be defined below: daa _ size + daa _ size inpu oupu T O = + laency (6 ne ( ij bandwidh The middleware ime cos is based on exising middleware. We simply specify he upper limi of he overall middleware cos T mid (O ij. Alhough we have described how o guaranee he nework bandwidh beween wo grid nodes and how o guaranee and specify he merics of T com (O ij and T mid (O ij in [9], T ne (O ij is difficul o specify as he daa size of services varies momenarily and he nework laency is difficul o guaranee. Here, in order o faciliae resolving he model, we assume he invariabiliy of daa_size inpu +daa_size oupu and specify he nework laency as a saisical feedback. (2 Repuaion Q ij.repuaion is a measure of he rusworhiness of O ij [11]. The value of he repuaion is defined as average ranking given o O ij by end users: ( O w k = 1 Rk ij Q. repuaion = (7 ij w where R k (O ij is he k h end user s ranking on he repuaion of O ij and w is he number of imes he service has been graded. (3 Successful execuion rae As defined in our previous work [1], Q ij.suc_rae is defined as he imes of successful O ij invocaions in proporion o he oal imes of O ij invocaions: ( Oij ( Oij N success Q. suc_ rae = (8 ij N oal (4 Availabiliy Q ij.avalabiliy is he probabiliy ha O ij is accessible. I is defined as: MTTF ( O ij Q. availabiliy = (9 ij MTTR ( Oij + MTTF ( Oij where MTTF(O ij is he Mean Time o Failure for O ij and MTTR(O ij is he Mean Time o Repair for O ij. (5 Price Q ij.proce is defined as he price charged for O ij by he provider. 4.2 Muli-dimension QoS scale We assign a number o each dimension of he vecor, from 1 o 5. The numbers denoe duraion, repuaion, successful rae, availabiliy, and price in ern. Thus Q vs can be described as a hree-dimension marix: Q vs =( Q ijk; 1 i u,1 j v, 1 k 5 (1 According o mehod proposed in [11], we classify he qualiy merics ino hree ypes. The firs ype, including Q ij.duraion, is negaive, i.e. he higher he value, he lower he qualiy. Anoher ype is posiive merics, including Q ij.repuaion, Q ij.suc_rae, Q ij.availabiliy, i.e. he higher he value, he higher he qualiy. Paricularly we classify Q ij.price ino a separae ype. The mehod proposed in [11] is adoped o eliminae he conrary characerisics beween negaive merics and posiive merics. For negaive merics, values are scaled according o (11. For posiive merics, values are scaled according o (12. Paricularly, we scale he value of price according o (13. Here, i denoes he sequence number of he candidae operaion providers. ( max ( Q Q jk ijk if ( Qjk ( Qjk max ( min ( 1 if ( Qjk ( Qjk Q min ( Q ijk jk if ( Qjk ( Qjk max( min ( 1 if ( Qjk ( Qjk V( Qij = ijk jk jk V Q Q Q ( V Q Q Q = ijk jk jk Based on he expression of V(Q ijk, formula (14 is used o describe he max performance/price for he j h funcion of he h VS. The weigh vecor ω=(ω k ; 1 k 5, ω k =1 is provided by he user, where value of ω k expresses how much aenion should be paid for he k h qualiy meric. max ( Q Q Q min 1 1 ( Q j ij 4 ijk jk * ω + 1 k = 2 ωk (14 max ( Q 1 min ( Q 1 max ( Q min j j jk ( Qjk V ( Q j = Max 5 ij. max min max min = max min max min = Q ω * ij5 5 The QoS merics of our model include bu are no limied o he five ones. New merics can be flexibly added wih he model unchanged. 4.3 QoS-opimized service composiion model Based on he above descripion, we esablish he mahemaical model for opimizing he composiion plan. The main idea of he model is o opimize performance/price of he composiion plan under he precondiion ha he oal duraion and oal price of he composiion plan are guaraneed wihin he limi given by he user. The mahemaical model is described as below: (11 (12 = Q (13 price
Given condiion (1~(4 (1 The direced acyclic graph G(E, V for he execuion pah has been generaed. Here V sands for he operaion se V={O 1, O 2,, O n } involved in he execuion pah and E is he se of he direced edges, which denoe he relaion among he operaions; (2 ( VS Grid, he qualiy marix Q vs =( Q ijk; 1 i u,1 j v, 1 k 5 is known; (3 The user gives he duraion limi of he composiion service as T oal, and he price limi of he composiion service as P oal ; (4 The user gives he weigh vecor ω=(ω k ; 1 k 5, ω k =1. Objecive funcion and consrains We wan o find a soluion, a composiion plan X=(x 1, x 2,, x n, where x i is seleced from all candidae providers, wih he opimized QoS expressed as he objecive funcion (15 and wo consrains (16 and (17. Here, in objecive funcion (16, V is he node se of he criical pah abou ime of he direced graph decided by he soluion X, or node se of he criical pah of X for shor. Max n V Q x (15 { ( 1 ( ( i = i k = 1 n ( = 1 ( ' ( = j ( j oal ( j T X duraion x T x V i i oal (16 P X = price x P (17 4.4 Model resolving procedure We assume ha he composiion service is made up of n operaions and here are m candidae providers for each operaion, he scale of he composiion plan will be m n by exhausive enumeraion and he compuaion cos of searching opimized execuion plan mus be very high even in very small scale of n and m. The simulaed annealing algorihm [12] is a echnique ha has araced significan aenion suiable for opimizaion problems of large scale. As T oal and P oal in he model are possible o be saisfied by a local exreme, we use simulaed annealing algorihm o opimize he oal duraion and price of he composiion plan. When consrains of (16 and (17 have been saisfied, we ry o opimize i = n 1 V( Q( xi using he local search algorihm [13] which is proven o be very useful o ackle opimizing problems arising from pracice. 4.4.1 Opiomizing T(X and P(X Using simulaed annealing algorihm, we firs ry o find a soluion o mee consrain (16, and hen adjus he soluion o saisfy consrain (17. Opimizing T(X from sep (1 o (3. (1 X=(x 1, x 2,, x n is a random plan of he composiion pah, where x i is randomly seleced from all he candidae providers. S=S is he sep couner. T=T is he iniial emperaure variable for he simulaed annealing algorihm. I decreases =T /(S *n each sep. Success= is a Boolean variable o indicae wheher he opimizaion has been successful. k X = ( x, x, L x k is he sequence of nodes in he 1 k2 k m criical pah of X abou execuion ime. (2 If S<, go o (3. Else, go on. Randomly selec a node x k j from X k. Replace x k j wih x k j ha has he minimum duraion among all he operaion providers offering he idenical funcion. We adjus X, and ge X ( * x,,, 1, x x x x k1 kj km n k ' Then we can ge X ( x ', x ',, x ' k1 k2 k p pah of X *. k ' If p ( = 1 ( ' p T X = duraion x T, i k oal L L L L. = L, he criical * X X, Success 1 and go o (3. Else, T T and accep X * (acceping X * means ( k ' k T X T( X * X X a probabiliy r = e T, S S 1, and go o he beginning of (2. (3 If Success=, T T and go o (1. Else Success, S S, T T and go o (4. Opimizing P(X from (4 o (5 (4 If S<, go o (5. Else, go on. Randomly selec a node x i from X=(x 1, x 2,, x n. Replace x i wih x * i, which has he minimum price among all candidae providers ha do no lead he oal execuion ime of he criical pah of X o surpass T oal. Then we ge he adjused X ( x, x, Lx, L x 1 2 i n. P X = n price x P, If ( = 1 ( i i oal Success, and go o (5. and accep X * a probabiliy X X, 1 Else, T T PX ( PX ( r = e T, S S 1 T T and go o (4. (5 If Success=, and go o (1. Else reurn X as he iniial value for opimizing Q(X. 4.4.2 Opimizing Q(X Using local searching algorihm, we sar from he iniial soluion of he composiion plan go in secion
4.4.1, and hen move ieraively hrough he soluion se. We make in each sep locally he bes choice, and sop if his does no lead o any improvemen. The algorihm for opimizing Q(X in described as below. According o sep (1~(5 we have X=(x 1, x 2, x i,, x n which saisfies (16 and (17. Given m i is he number of he candidae providers of x i, we define N i (X as a se of alernaive soluions of X for x i Ni ( X = {( x, LL, x,, x, L, ( x,,,,, (,,,,, (, } 1 i 1 n 1L x L x x L x L x x x x i k 1 n 1 i k+ 1 n LLL 1 i m n We ge he subse of N i (X, i.e. N * (x i, he elemens of which keep consrains (16 and (17 saisfied. The Neighborhood Srucure [13] of X is defined p as N ( X = U N ( x k i, where xk i = 1 i criical pah of X is k X ( x, x, x k1 k2 k p = L. k X, and he In he firs sep, only he nodes on he criical pah are aken ino accoun for efficiency of he algorihm. This consideraion is based on he following proven heorem. Theorem1. Given G (E, V is a Direced Acyclic Graph (DAG, V is he se of verexes on he Criical Pah (CP of G(E,V abou Time. ( x V ', if he ime value of x decreases, he oal ime of G(E,V will no increase; if he ime value of x increases, he oal ime of G(E,V will increase oo and he CP will no change. From he curren soluion X, he algorihm is o p N X = U N x for X * ha generaes search ( ( k 1 i ( i = Max n 1V ( Q ( xi. If ( i = Max n 1V ( Q ( xi is larger han he oal qualiy of soluion X, replace X wih X * (i.e. X X. Repea searching new X * in he same way, unil i = Max n 1V ( Q ( xi does no increase. However, ( insead of sopping he algorihm here, we apply he following procedure o opimize he qualiy of noncriical-pah nodes. The main idea o opimize he qualiy of he noncriical-pah nodes is based on Theorem 2. From he non-criical-pah nodes we uniformly selec one every ime and replace i wih he candidae provider ha generaes he maximum qualiy value under he precondiion ha he execuion duraion of he candidae provider for his node does no surpass he execuion ime of he curren provider of his node and he oal price of he new soluion does no surpass P oal. Repea selecing and opimizing he remaining noncriical-pah nodes unil all he non-criical-pah nodes have been hi. Theorem2. Given G (E, V is a Direced Acyclic Graph, V is he se of verexes on he Criical Pah (CP of G (E, V abou Time. ( x ( V V ', if he ime value of x does no increase, he CP will no change. 5. Simulaion and Experimenal Resuls In his secion we analyze he performance and efficiency of our soluion o he model proposed in secion 4. In he simulaion, he simulaor randomly generaes DAGs wih differen number of verexes for he execuion pah of he composiion service. For each node in a DAG, he simulaor generaes qualiy marixes o express he QoS merics from differen providers. We conrol he scale of he opimizaion by changing he number of he operaions involved in he composiion service n and he number of candidae providers for each operaion m from 1 o 1. 5.1 Compuaion ime for qualiy opimizaion in differen scale In his experimen, he simulaor opimizes he qualiy of he composiion service of every given scale, for example, n=1 and m=3, 4 imes, and saisically ges he average compuaion ime for each scale. I can be seen from Figure 6 ha alhough he compuaion cos increases wih he number of he elemens of he composiion plan more and more quickly, he opimizing ime cos of a pracical range of scale is accepable. For example, when he composiion service is made up of 1 aomic operaions and 3 candidae providers are available for every operaion, he compuaion ime for qualiy opimizaion is only 498 milliseconds. Figure 7 shows ha when he number of elemen of he composiion service is fixed, he oal compuaion ime for opimizaion has a near linear relaionship wih he number of he candidae providers. T: compuaion ime cos for qualiy opimizaion(in milli-seconds 5 45 4 35 3 25 2 15 1 5 Performance Changing wih he Number of Operaions Involved Number of Providers= 5 Number of Providers= 1 Number of Providers= 2 Number of Providers=3 1 2 3 4 5 6 7 8 9 1 N: number of operaions involved in he composiion service Figure 6. Performance changing wih n
1 2 3 4 5 6 7 8 9 1 T: compuaion ime cos for opimizaion(in milli-seconds Q: oal qualiy value of he composion plan Q: oal qualiy value of he composiion plan 45 4 35 3 25 2 15 1 5 3 25 2 15 1 5 Performance Changing wih he Scale of Candidae Providers Number of Elemens of he Composiion=1 Number of Elemens of he Composiio=2 Number of Elemens of he Composiio=3 Number of Elemens of he Composiio=4 M: number of candidae providers for each operaion Figure 7. Performance changing wih m Qualiy Value in Differen Phases of he Opimizaion Iniial Qualiy Value Qualiy Value afer Opimizaion Phase1 Qualiy Value afer Opimizaion Phase2 Qualiy Value afer Toal and Poal Consrain Saisfied 1 2 3 4 5 6 7 8 9 1 N: number of operaions involeved in he composiion service Figure 8. Qualiy in differen phases (m=3 14 12 1 8 6 4 2 Qualiy Value in Differen Phases of he Opimizaion Iniial Qualiy Value Qualiy Value afer Toal and Poal Consrain Me Qualiy Value afer Opimizaion Phase1 Qualiy Value afer Opimizaion Phase2 1 2 3 4 5 6 7 8 9 1 M: number of candidae providers of each operaion Figure 9. Qualiy in differen phases (n=4 5.2 Effec of he opimizaion algorihm In his experimen, we ry o analyze he effec of our algorihm for solving he model. Based on he above descripion of he algorihm, we pariion he procedure of opimizing qualiy ino 4 phases, he random selec composiion plan, he adjused composiion plan afer (16 and (17 have been saisfied, he opimized composiion plan afer he qualiy on he criical pah has been opimized, opimizaion phase 1, and he opimized composiion plan afer he non-criical-pah nodes are opimized, opimizaion phase 2. The simulaor saisically figures ou he average qualiy value improved in each phase. In his experimen he weigh vecor is se as ω=(ω k =.2; 1 k 5. The emperaure variable for he simulaed annealing mehod is se as T=1, and i decreases =T/(S *n each sep. Here, he sep lengh is se as S =5. The qualiy vecor of each candidae operaion provider is randomly generaed in he following consrains. 15 Q ij.duraion 3, 1 Q ij.repuaion 1, 5% Q ij.suc_rae 1%, 5% Q ij.availabiliy 1%, 1 Q ij.price 3. Figure 8 shows ha when he number of provider for each operaion is fixed, he oal qualiy of he composiion plan is grealy improved especially in opimizaion phases 1 and opimizaion phase 2. Figure 9 shows ha when he number of operaions is fixed, he final opimized qualiy of he composiion plan will harves more improvemen if here are more providers for each operaion. In some cases he oal qualiy may decrease in he second phase. The reason is ha he iniial soluion has beer qualiy bu consrain (16 or (17 may be unsaisfied and he qualiy of he composiion plan mus be adjused. 6. Relaed Works QoS problem arises from resource compeiion among applicaions. Much research work has been done for resource reservaion [14] and service level agreemen [15] in grid environmen. Auomaic service composiion is a very acive area of research and sandardizaion. In [16], a web service composiion oolki, SWORD, was developed o use a rule-based exper sysem o auomaically generae a composie service. However, hey give no deails abou he exper sysem, and we are inspired o design he rule-based algorihms o generae he composiion plan. Table 1. Comparision of Agflow and Q-SAC Agflow Q-SAC Mehod Experimen configuraion Performance n=4, m=4 Middleware Cluser of PCs IBM s OSL (Penium III 933MHz, seconds 512M RAM, Windows 2 Mahemaical PC(Inel Celeron 4A, model simulaion 2.GHz, 256M RAM, seconds Windows 23 T opimize=1.6 T opimize=.11 Very limied work in he QoS opimizaion of composie service has been specifically addressed. AgFlow [12] addresses he issue of selecing Web Services for he purpose of composiion. IBM s Opimizaion Soluions and Library are used o implemen he global planning. Unlike heir work, we esablish a perfomance/price oriened mahemaical model for QoS opimizaion of composiion plan and design an algorihm o solve he model. Alhough more
deail is needed o compare he wo works, Table 1 gives some comparison beween our simulaion and he experimen on heir middleware. 7. Conclusions and Fuure Work In his paper, Q-SAC is presened for opimizing QoS of he auomaically generaed composie grid services. We design a rule-based composiion algorihm o generae he composiion pah of he composie service and esablish a mahemaical model for opimizing he QoS of he global composiion plan. Algorihms are designed for solving he mahemaical model. Simulaion resuls show ha our model and algorihms are efficien. The inpu daa and oher deail informaion for he simulaor can be accessed a websie hp://grid.hus.edu.cn/hhchen/sim/qsac. In nex seps, we will analyze and improve he precision of he global QoS value opimized by Q-SAC and comparing wih oher ime-consuming bu precise algorihm. References [1] I. Foser, C. Kesselman, J. M. Nick, and S. Tuecke, Grid Service for Disribued Sysem Inegraion, IEEE Compuer, Vol.35, No.6, June 22, pp.37-46. [2] I. Foser, C. Kesselman, J. M. Nick, and S. Tuecke, The Physiology of he Grid An Open Grid Services Archiecure for Disribued Sysems Inegraion, DRAFT documen of Globus Projec of he Universiy of Chicago, February 17, 22. [3] H. Jin, ChinaGrid: Making Grid Compuing a Realiy, Digial Libraries: Inernaional Collaboraion and Cross-Ferilizaion, LNCS, Vol.3334, Springer-Verlag, 24, pp.13-24. [4] K. Czajkowski, D. F. Ferguson, I. Foser, J. Frey, S. Graham, I. Sedukhin, D. Snelling, S. Tuecke, and W. Vambenepe, The WS-Resource Framework, hp://www.globus.org/wsrf/. [5] J. B. Weissman and B. D. Lee, The Service Grid: Supporing Scalable Heerogeneous Services in Wide- Area Neworks, Proceedings of 21 Symposium on Applicaions and he Inerne (SAINT 21, San Diego, CA, Jan. 21, pp.95-12. [6] S. A. McIlraih, T. C. Son, and H. L. Zeng, Semanic Web Services, IEEE Inelligen Sysems, Vol.16, No.2, 21, pp.46-53. [7] O. Papini, Knowledge-base revision, The Knowledge Engineering Review, Vol.15, Cambridge Universiy Press, UK, pp.339-37. [8] V. Tosic, B. Esfandiari, B. Paguredk, and K. Pael, On requiremens for Onologies in Managemen of Web Services, Web Services, E-Business, and he Semanic Web, LNCS, Vol.2512, Springer-Verlag, pp.237-247. [9] H. Jin, H. Chen, J. Chen, P. Kuang, L. Qi, and D. Zou, Real-ime Sraegy and Pracice in Service Grid, Proceedings of The Tweny-Eighh Annual Inernaional Compuer Sofware & Applicaions Conference (CompSac 4, IEEE Compuer Sociey, Hong Kong, China, Sepember 28-3, 24, pp.161-166. [1] H. Chen, H. Jin, M. Zhang, P. Tan, D. Zou, and P. Yuan, Early Experience in QoS-Based Service Grid Archiecure, Advanced Web Technologies and Applicaions, LNCS, Vol.37, Springer-Verlag, March 24, pp.924-927. [11] L. Z. Zeng, B. Benaallah, A. H. H. Ngu, M. Dumas, J. Kalagnanam, and H. Chang, QoS-Aware Middleware for Web Services Composiion, IEEE Transacions on sofware engineering, Vol.3, No.5, May 24, pp.311-327. [12] R. H. J. M. Oen and L. P. P. P. van Ginneken, The annealing algorihm, Kluwer Academic Publishers, 1989. [13] E. H. L. Aars and J. K. Lensra, Local search in combinaorial opimizaion, John Wiley & Sons, Chicheser, 1997. [14] B. Urgaonkar and P. Shenoy, Sharc: Managing CPU and Nework Bandwidh in Shared Clusers, IEEE Transacions on Parallel and Disribued Sysems, Vol.15, No.1, Jan, 24, pp.2-17. [15] K. Czajkowski, I. Foser, C. Kesselman, V. Sander, and S. Tuecke, SNAP: A Proocol for negoiaing service level agreemens and coordinaing resource managemen in disribued sysems, Job Scheduling Sraegies for Parallel Processing, LNCS, Vol.2537, Springer-Verlag, 22, pp.153-183. [16] S. R. Ponnekani and A. Fox, SWORD: A Developer Toolki for Web Service Composiion, Proceedings of WWW, 22.