1 Interntionl Journl of Computers nd Applictions, Vol. 9, No., 007 WEB DELAY ANALYSIS AND REDUCTION BY USING LOAD BALANCING OF A DNS-BASED WEB SERVER CLUSTER Y.W. Bi nd Y.C. Wu Abstrct Bsed on our survey of recent rticles, there is little reserch being conducted into quntittive nlysis of the lod blncing of Web server clusters. In this pper, we propose quntittive nlysis for DNS-bsed server clusters. We lso propose two-pss lod-blncing method for determining the lod blnce re of these clusters. The first pss uses the lookup tble insted of complicted computtion for obtining the lod blncing of the Web service requests. The second pss lso utilizes lookup tble by precomputed Hessin mtrix to obtin the lod blncing. In ddition, we compre the reltive performnce of disptcherbsed nd DNS-bsed server clusters using queuing theory, nlysis, nd simultion, nd we compre the mesurement results using benchmrks. To increse the simultion performnce we hve designed simultion module to promptly locte the lod blncing, with potentil improvement of 36.58% over the verge system response time. Key Words System performnce, modelling techniques, mesurement techniques, web servers, lod blnce, system response time. Introduction The informtion trffic on the world wide web is incresing t the rte of bout two or three times over per yer nd the verge login-time on user s web pge is much longer thn it ws before. Web dely or system response time is one of the min fctors determining the qulity of service of web pge, nd networkbndwidth nd connection strtegy cn lso ffect the totl ltency sources. Mny mens of reducing web dely hve been studies, such s cching rchitectures, prefetching, content distribution network, trnsmission scheme, efficient content, imge optimiztion, nd lod blncing [ 4]. Lod-blncing techniques cn be ctegorized into four mjor clsses: client-bsed pproch, disptcher-bsed p- Deprtment of Electronic Engineering, Fu Jen Ctholic University, Tipei, Tiwn, 4, R.O.C.; e-mil: Recommended by Dr. Weimin Zheng (pper no ) 79 proch, DNS-bsed pproch, nd server-bsed pproch. Most of these investigtions provide new design of the rchitecture nd the lgorithm to gin performnce improvement. However, becuse not much is known bout the quntittive nlysis of vrious rchitectures, we propose the corresponding queuing model nd prompt method to determine the lod blncing of DNS-bsed web server clusters [4, 5]. Due to the stochstic chrcteristics of web server opertion nd internet trnsmission, web dely cn be very rndom. Therefore it cn be difficult to estimte the precise dely. Our reserch includes finding the key fctors utilized in estimting the verge system response time of the DNS-bsed pproch nd using one of these lodblncing pproches to reduce web dely . Bsed on the lod-blncing mechnism of DNS-bsed pproch, we provide mthemticl model, to nlyze nd simulte results, mesure the verge system response time of web server cluster, nd find some of the dependent fctors [ 5]. First we choose DNS-bsed web server cluster becuse it is commonly used nd hs not been studied in ny previous quntittive nlysis. For tht web server cluster we hve designed queuing model to conduct quntittive nlysis nd to find mthemticl reltion to the web dely of vrious disptch-fctors. Bsed on these results, one cn djust the system opertion prmeters to obtin the minimum possible web dely. Second, we use the mthemticl softwre pckge QNAT (queue networknlysis tool) [6 8] to conduct simultion. In ddition, we propose mthemticl model bsed on queuing model to compute either the web dely or the verge system response time E(t). The mthemticl model will let us know the chrcteristics of the web dely E(t). Then we use the QNAT nd TK Solver softwre pckges to verify whether our mthemticl model is correct. Furthermore, we hve designed new simultion tool to reduce the simultion time by finding the lodblncing condition of cluster of web servers. We lso use modified version of the Hessin mtrix, H[E(t)], to show how we locte the opertion region of the lod blncing of cluster of web servers. Finlly, we provide precomputtion tble to determine how to find lod blncing without using excessively complicted computtion.
2 The rest of this pper is orgnized s follows. In Section the min fctors of web dely re described. In Section 3 severl lod-blncing schemes re given nd resons for choosing the DNS-bsed rchitecture re presented. Section 4 describes the queuing model. Section 5 describes the lod-blncing scheme of cluster of web servers on DNS cluster rchitecture with Hessin mtrix utilizing two-pss lod blncing. Section 6 provides the performnce comprison between disptcher-bsed nd DNS-bsed rchitectures. Section 7 shows the experimentl results of the verge system response time with lod blncing using DNS-bsed web server cluster. Some conclusions from our reserch re drwn in Section 8.. The Min Fctors of Web Dely. Ltency Sources of Systems When we use web browser or ccess version of windows, series of computer system nd networkprocesses tkes plce in the bckground. Every process my hve rndom dely time of bout millisecond to second, depending on different opertionl situtions. For exmple, in clientserver system one first sends request to the browser in the website, where the browser will service the ccepted request. Most of the web contents must wit for other module responses such s softwre or hrdwre. In the best situtions time is required to execute their functions no mtter whether limited by the service bndwidth of the link or by device services. So ltency sources cn be the cuse of most inherent delys nd cn mke it difficult to improve the performnce of common computer networks [ 3].. DNS Lookup Tble When we go to website we must remember the domin nme, such s In typicl computer networkit is necessry to trnsform the domin nme to its specific IP ddress. We lso find tht 0% of the domin nme trnsformtion tkes bout 6 seconds due to retrnsfer time. As the number of clients increses, the DNS lookup tble dely time becomes longer. So the DNS lookup is lso prime fctor cusing web dely [ 3]..3 Connection Therefore the server frequently becomes bottleneck, nd server-side processing lso ffects the bndwidth tht web session needs [ 3]..5 Document Trnsfer Depending on content size, vilble bndwidth, setting of proxy servers nd routing, document trnsfer will be mjor fctor cusing dely time. For exmple, the trnsfer time of MB of dt will be longer thn tht of KB of dt for the sme trnsmission rte. Therefore the document trnsfer time cn be n uncontrollble fctor of dely . 3. Previously Proposed Approches 3. Client-Bsed Approch Client-bsed pproch routes the request to one of the destintion nodes on the web clusters to specific server bsed on the client-site softwre. This pproch is divided into web client nd client-side proxies. The stte mngement of the client-bsed pproch is very difficult when network trffic increses. As the client-bsed pproch is not often used, we will not consider using it [ 3]. 3. DNS-Bsed Approch Due to the increse of web trffic nd longer URL nmes, we hope to mke simple virtul interfce to communicte with others. This interfce will mke it esier for users who re not experts. A DNS-bsed pproch just trnsforms the site nme of the distributed nodes of the web system into n IP ddress. So, we must design solution with couple of steps s shown in Fig. in order to distribute the requests of ll clients to the most suitble server. Of course, mny intermedite nme servers cn cche logiclto-ip-ddress mpping to reduce the mount of network trffic between client nd the cluster DNS. Therefore DNS-bsed TTL (time-to-live) cn be used to determine how much time will be required to find n intermedite nme server; otherwise the DNS cluster cnnot select suitble web server. This pproch is good ide, but the cluster DNS server often cuses dditionl delys, s too much trffic psses through the intermedite nme server [ 7]. The next web dely builds up the TCP to route the web service to the destintion bsed on the IP dtgrm. A routing lgorithm provides solution to determine how to trnsfer the service request to the next stge or to the wireless-level. So router lso contributes to the dely time. We further found tht bout 40% of the connections tke bout 00 0,000 milliseconds [ 3]..4 Server-Side Processing Due to the increse of websites nd new services, the working lod of web servers increses, even for personl business ffirs including dynmic content. The dynmic dt need to mke more I/O nd CPU requests to the servers. 80 Figure. DNS-bsed pproch rchitecture.
3 3.3 Disptcher-Bsed Approch To centrlize the scheduling request nd control clientrequest routing completely, disptcher-bsed pproch hs been designed. The routing request mong servers is trnsprent, but this is the sme when deling with ddresses t the URL level on DNS. A typicl disptcher hs simple virtul IP ddress (IP-SVA). The disptcher defines its own personl ddress bsed on its server nd its different protocol levels bsed on distinct structures such s pcket rewriting, pcket forwrding, or HTTP redirection. The disptcher selects n lgorithm to choose specific web server to blnce incoming requests to minimize processing dely, s shown in Fig. [ 3, 9 ]. Figure. Disptcher-bsed pproch rchitecture. 3.4 Server-Bsed Approch The server-bsed pproch uses two-level system of disptching. This pproch first distributes the client requests on the web DNS to the web-server nodes. Then every server distributes the request to other system servers gin. Of course, its dvntge is to conduct lod blncing more effectively. In contrst, implements nd mngements re more inconvenient [, 4]. 4. Queuing Models In rel situtions pckets will be lost due to collision or lckof sufficient buffer. Collision is cused by more thn two pckets being trnsferred t the sme time on the MAC lyer. Lckof sufficient buffer is often cused by the service rte being less thn the rrivl rte. To observe the effect of lckof sufficient buffer we ssume tht the MAC lyer is idel in our pper. This mens pckets will only be lost by lckof sufficient buffer, not by collision. There is one intermedite nme server, two dns servers, nd two web servers in our rchitecture. To provide quntittive nlysis of the system performnce, we propose corresponding queuing model for the DNS-bsed rchitecture. In this model there re two clsses of rrivl rtes, one for the ddress request nd the other for the document request. The domin nme server provides service for the ddress request, nd then the client sends the document request to the web server designted by the domin nme server. This system is modelled s 8 shown in Fig. 3, nd the system prmeters re shown in the following. Figure 3. DNS-bsed queuing model. Due to the simultion of networkdely, we dd queue t the feedbckpth s shown in Fig. 3 with feedbck rtio of i nd j to pproximte the rel sitution. The prmeters i nd j re the DNS Server nd DNS Server pcket loss rtios, respectively. For exmple, if i is 0% nd λ is 00 (pckets), 0 pckets will be feedbck to the intermedite nme server from the DNS server to retrnsfer. Definition of system prmeters: λ : Address request rte (job/sec) λ : Document request rte (job/sec) µ 0 : Intermedite nme server service rte (job/sec) µ : Cluster domin nme server service rte (job/sec) µ : Web server service rte (job/sec) : Disptch rtio, 0 < i, j: Feedbckrtio, 0 <i,j λ kλ E (t): The front-end of the domin nme server response time (sec) E (t): The bck-end of the domin nme server response time (sec) E(t): DNS-bsed system response time (sec), E(t) E(n)/λ E(n)ρ n /( ρ n ). The verge number of tsks or jobs on queue To simplify the simultion, we use n infinite length of the buffer so tht there will be no blocking probbility. In ddition, the methods of TLL re round robin. 4. QNAT Simultion Usully, there is three-performnce index for the comprison of simultion results. Here we use the system response time E(t) s the comprison index for the simultion. There re three cses: without networkdely, with networkdely, nd with vrition feedbcknd network dely. Among the three we find wy to locte the lod blncing by using the friendly queuing simultion tool QNAT [, 6 8]. To simulte the performnce we need to ssign specific set of prmeters, n intermedite nme server service rte of 00 (job/sec), cluster DNS service rte if (job/sec), nd web server service rte of 500 (job/sec), ρ.
4 4. Without Network Dely In the cse of without networkdely we just ignore the networkdely. From Tble nd Fig. 4 of the simultion results we find tht there is lod blnce nd minimum system response time when the disptch rtio 0.5, becuse the service rtes of servers nd in Fig. 4 re equl. Tble Front-End System Response Time of DNS-Bsed Architecture (without NetworkDely), µ 0 µ µ 00 (job/sec), i j 0. Arrivl Rte Disptch Rtio Front-End System λ (job/sec) Response Time E (t) (sec) shows the front-end system response time with vrible feedbcknd networkdely. Figure 5. The front-end system response time of DNSbsed rchitecture (with networkdely), µ 0 µ µ 00 (job/sec), i j 0.. Figure 4. Front-end system response time of DNS-bsed rchitecture (without networkdely), µ 0 µ µ 00 (job/sec), i j 0., λ 0 (job/sec). 4.3 With Network Dely Becuse there is dely in networktrnsmission, we should dd on the networkdely to the simultion. From Fig. 5 of the simultion results we see tht there is lso lod blnce for the disptch rtio of 0.5, no mtter wht the networkdely is in the model, s shown in Fig. 3. Fig. 8 Figure 6. The front-end system response time E (t) (sec) with vrible feedbcknd networkdely λ 0 (job/sec), µ 00 (job/sec), µ increses 0 (job/sec), nd i, j reduce Mthemticl Models Bsed on Fig. 3 we derive E (t), E (t), nd E(t) s shown in (3), (4), nd (5). The derivtion cn be seen in . The bsic steps re: finding the utiliztion ρ of ech queue, finding the verge number of tsks or jobs E(n)ρ n /( ρ n ) on ech queue, nd then finding the verge time E(t) E(n)/λ for single tskto get through the queue. Hence we obtin the following equtions () through (4), where W q represents the network dely. E (t) ( i j j)µ λ ( i j j)µ λ W q ( i j j)µ ( )λ ()
5 W q ρ λ ρ i ( i j j)µ iλ ( )j () ( i j j)µ ( )jλ E (t) ( i j j)µ λ ( i j j)µ λ ( i j j)µ ( )λ i ( i j j)µ iλ ( )j (3) ( i j j)µ ( )jλ E (t) (4) µ λ µ kλ If we hve the sme service rte of the DNS server cluster nd very smll networkdely, then we cn neglect W q. Hence we obtin the verge system response time E(t) s follows: E(t) ( i j j)µ λ ( i j j)µ λ ( i j j)µ ( )λ i (5) ( i j j)µ iλ µ kλ 5. Lod Blncing on Cluster DNS Using Hessin Mtrix E(t) ( i j j)µ λ ( i j j)µ λ ( i j j)µ ( )λ i ( i j j)µ iλ (7) µ kλ 5. Two-Pss Lod Blncing DNS-bsed rchitecture needs to perform two-pss requests while providing web services. The first pss is the lod blncing of ddress request, wheres the second is the lod blncing of document request bsed on the distribution of ddress request. We find tht the service rte of the front end will ffect the feedbckrtio nd the document request rte. If the feedbckrtio of the ddress request increses nd the lod blncing of the DNS side is unstisfctory, the document request rte decreses. In other words, if the lod blncing of the DNS side is sufficient, the feedbck rtio is reduced nd the document request rte increses. Therefore, ccording to the service request rte, we djust the DNS service rte nd locte the minimum system response time. The procedure is shown in Fig. 7. Here we will define the reltionship between k nd the feedbck rtio j. To obtin simultion exmple, we provide specific set of prmeters. We set the service rrivl rte of the DNS cluster t 00 (job/sec) nd λ λ. From () we lern tht E (t) is similr to the model from the previous reserch for the disptcher-bsed rchitecture . There is one smll difference: the networkdely. Given specific set of prmeters, we estblish the fct tht the networkdely is (sec) for given specific set of prmeters s shown in (6). Hence we use the Hessin mtrix  to form precomputed tble insted of complicted computtion for finding the lod blncing of the system. From comprison of the simultion results we lso find fixed difference of (sec). Hence the system response time E(t) cn be represented by (7). W q ρ λ ρ ( ) ( 0.5) 0. ( ) ( 0.5) (6) 83 Figure 7. Two-pss flow chrt for locting the lod blncing of DNS-bsed rchitecture.
6 In ddition, when the service rte decreses every 0 (job/sec), k decreses by. Hence the vlue of k is locted in the intervl 6 becuse the service rte of the DNS cluster rnges from 40 to 00 (job/sec). As in the previous section, we lso ssume tht the feedbckrtio is reduced by step size of 0.0 when the service rte is incresed by step size of 0 (job/sec). Therefore we cn obtin (8) nd (9) s the reltionship of k, j nd service rte of the DNS cluster s shown in Fig. 3. k (0. j) 00 (8) µ 40(j 0.4) 00 (9) Becuse we use the sme set of cluster DNS, the bckend of web servers hs set of similr prmeters. Hence we cn rewrite E (t), E (t), nd E(t) s in (0) to (). E (t) ( j) 00 λ ( j)(40 (j 0.4) 00) 0.5λ (0) E (t) µ ( (0. j) 00)λ () E(t) ( j) 00 λ ( j)(40 (j 0.4) 00) 0.5λ () µ ( (0. j) 00)λ According to the mthemticl model nd the serch procedure for lod blncing shown in Fig. 7, we cn locte the lod blncing quickly. As we exmine Tble, we find tht when µ µ 90 there is minimum system response time E(t). According to Tble we reduce the system response time by ( )/ %. 5.3 Lod Blncing of the Whole System Using Hessin Mtrix Previously we proposed lookup tble method to locte the lod blncing for E (t). Here we extend single-pss method to two-pss lod blncing due to the dependence of E (t) nd E (t) on ech other. For both E (t) nd E (t) we provide first-pss mechnism tht uses the lookup tble method to locte the lod blncing nd then provide the second-pss lod blncing by using precomputed tble of the Hessin mtrix for document request [, 3]. First, we tke prtil derivtive of (5) with respect to j nd λ : E(t) 00 (00 00j λ ) 00j 74 ( 00j 74j 6 0.5λ ) 00λ (500 λ 00jλ ) (3) E(t) λ (00 00j λ ) 0.5 ( 00j 74j 6 0.5λ ) 4 00j (500 λ 00jλ ) (4) Second, we tke prtil derivtive of (3) nd (4) with respect to j nd λ : E(t)/ 0000 (00 00j λ ) j 44400j 65 00λ ( 00j 74j 6 0.5λ ) λ (500 λ 00jλ ) 3 (5) Tble System Response Time of DNS-Bsed Architecture λ 0 (job/sec), µ (job/sec), µ increse 0 (job/sec), nd i, j reduce 0.0 Disptch Service Rte of System Response System Response System Response Rtio DNS Cluster time of Front-End Time of Bck-End of Time of DNS-Bsed µ (job/sec) of DNS-Bsed DNS-Bsed Architecture (E(t)) Architecture (E (t)) Architecture (E (t))
7 E(t)/ 00 λ (00 00j λ ) 3 00j 74 ( 00j 74j 6 0.5λ ) λ 0000jλ (500 λ 00jλ ) 3 (6) E(t)/ λ λ (00 00j λ ) ( 00j 74j 6 0.5λ ) 3 E(t)/ λ 6 600j 40000j (500 λ 00jλ ) 3 (7) 00 (00 00j λ ) 3 00j 74 ( 00j 74j 6 0.5λ ) λ 400jλ (500 λ 00jλ ) 3 (8) Third, we form Hessin mtrix: E(t)/ E(t)/ λ E(t)/ λ E(t)/ λ λ (9) To locte the minimum system response time E(t) from (9) we must let the determinnts of (9), nd, be greter thn zero. If j 0.4, then: E(t)/ 0000 ( λ ) (0.4) (0.4) 65 00λ ( 00 (0.4) 74 (0.4) 6 0.5λ ) λ (500 λ 00 (0.4)λ ) 3 (0) E(t)/ 00 (00 00 (0.4) λ ) 3 00 (0.4) 74 ( 00 (0.4) 74 (0.4) 6 0.5λ ) λ 0000 (0.4)λ (500 λ 00 (0.4)λ ) 3 () E(t)/ λ λ ( λ ) 3 E(t)/ λ 0.5 ( 00 (0.4) 74 (0.4) 6 0.5λ ) (0.4) (0.4) (500 λ 00 (0.4)λ ) 3 () 00 ( λ ) ( 00 (0.4) λ ) λ λ (500 λ λ ) 3 (3) From the existing criteri of the minimum vlues of Hessin mtrix, we hve: ( E(t)/ E(t)/ λ λ E(t)/ E(t)/ λ ) > 0 λ (4) We find the solution intervl 9.4 λ < 3.5 from the intersection of the solutions from (0) to (4). In ddition we obtin both the j locted in the intervl [0.4 0.] nd the reltionship between j, λ, nd µ ; then we obtin similr results from (0) to (4) s shown in Tble Comprison of Disptcher-Bsed nd DNS- Bsed Architecture In the simultion results with specific set of prmeters shown in Fig. 8, where the rrivl rte becomes 80 (job/sec), the system response time of disptcher-bsed Tble 3 Two-Pss Lookup Tble Arrivl Rte λ (job/sec) Service Rte of Cluster DNS µ (job/sec) System Response Time of DNS-Bsed Architecture 9.4 λ < Hs minimum system response time 7.77 λ < Hs minimum system response time 6.3 λ < Hs minimum system response time 5 λ < Hs minimum system response time 3.8 λ < 5 80 Hs minimum system response time.73 λ < Hs minimum system response time 85
8 rchitecture my be very long s compred to the DNSbsed rchitecture . However, when the rrivl rte is smll, the system response time of disptcher-bsed rchitecture will be very short. Furthermore, when the rrivl rte becomes 63 (job/sec), the system response time of disptcher-bsed nd DNS-bsed rchitecture will be equl. Hence, for the smll rrivl rte in LAN, we will choose disptcher-bsed rchitecture insted of DNS-bsed rchitecture. In other words, for the lrge rrivl rte in n Internet, we will choose DNS-bsed rchitecture insted of disptcher-bsed rchitecture, lthough DNS-bsed rchitecture hs higher mngement cost . 7. Mesurement of Lod Blncing Using DNS- Bsed Web Server Cluster To verify the chrcteristic of the proposed model s shown in Fig. 3, we designed nd implemented testing system with lod-blncing mechnism of DNS-bsed web server cluster s shown in Fig. 9. In ddition, we use the mesurement tool webserver stress tool stndrd edition to mesure the system response time of the web requests. Tble 4 shows the system specifiction . First, we must set the prmeters for the testing system s shown in Tble 5 with nd without lod-blncing mechnism. Second, by mesurement, we obtin the system response time with respect to number of times s shown in Fig. 0, which shows the comprison between with lod blncing nd without lod blncing setting. We lso lern from the specific mesurement results nd (5) tht we hve potentil improvement of 76% with lodblncing mechnism. Tble 5 Prmeter Setting of Webserver Stress Tool Figure 8. A comprison of lod blncing between disptcher-bsed pproch nd DNS-bsed pproch. Number of Mesured Mx Dt Number of Web Users Time Rte Servers 0 40 Minute 64 kb Figure 9. Testing system for the DNS-bsed web server cluster. Tble 4 System Specifiction of Web Servers CPU RAM (Mb) O.S. Network(Mb/sec) Intermedite Nme Server One P4.6 GMHZ 768 Windows 000 Server 00 DNS One P4.6 GMHZ 768 Windows 000 Server 00 DNS One P4.6 GMHZ 768 Windows 000 Server 00 Web Server One P4.4 GMHZ 5 Windows 000 Server 00 Web Server One P4.4 GMHZ 5 Windows 000 Server 00 86
9 in Fig., which shows both results, from the G/M/ mesurement curve nd from the M/M/ model curve. The increse in the number of users increses the system response time becuse it cuses the server system to be busier, thus delying the service requests. The G/M/ mesurement curve provides the reltionship between the system response times with respect to the generl rrivl distribution. Figure 0. Comprison of the system response time between with lod blncing nd without lod blncing % (5) Third, by mesurement, we lso obtin the system response time with respect to the disptch rtio s shown in Fig., which shows tht djustment of the disptch rtio cn ffect the system response time. Becuse servers nd in Fig. 9 re t the sme mchine model they hve the sme service rte. Therefore, s the disptch rtio is 0.5, which is evenly distributed s shown in Fig., we find tht here we hve the minimum system response time by mesurement. These results of lod blncing re lso similr to the simultion results from QNAT shown in the previous sections. Figure. System response time of testing system of DNS-bsed rchitecture. Fourth, by mesurement, we lso obtin the system response time with respect to number of users s shown 87 Figure. Comprison of the system response times of M/M/ model nd G/M/ mesurement. Bsed on the queuing theory, the M/M/ model of the system response time of DNS-bsed rchitecture is shown in (6) . E (t) ( i j j)µ λ ( i j j)µ λ ( i j j)µ ( )λ i (6) ( i j j)µ iλ Bsed on the queuing theory the G/M/ model of the system response time of DNS-bsed rchitecture is shown in (7) . [ E (t) ( i j j)µ λ ( i j j)µ λ ( i j j)µ ( )λ ]( i C ) (S) ( i j j)µ iλ (7) C (S) in (7) is vrition coefficient. In the M/M/ model C (S) is equl to, wheres in our G/M/ mesurement C (S) is lrger thn s shown in Fig. if the
10 vrition of the system response time is lrger thn the men of the system response time [, 8]. The opertionl chrcteristics of the testing system mke the vrition coefficient shift s shown in Fig.. Usully the vrition coefficient of the testing system is dependent on the networksitution, the system softwre mngement policy, nd the chrcteristics of the mesurement tool. Hence couple of unknown fctors my ffect the mesurement results. Although the G/M/ mesurement curve is slightly fr from the M/M/ model curve, due to the lrge vrition coefficient in the G/M/ mesurement both curves still hve similr trend, s shown in Fig.. 8. Conclusion In this pper we propose two-pss lod-blncing scheme. The first pss is for the ddress request nd the second is for document request. Bsed on the corresponding queuing model for DNS-bsed rchitecture, we cn nlyze nd simulte the system response time. Furthermore, by utilizing the precomputed tble of Hessin mtrix insted of complicted computtion, we cn esily locte the lod blncing. Finlly, from comprison of the system response times between DNS-bsed nd disptcher-bsed rchitecture we lern tht disptcher-bsed rchitecture is good for LAN pplictions nd DNS-bsed rchitecture is good for internet pplictions. From nlysis, simultion, nd mesurement results we lern tht when lod blncing exists due to specific set of prmeters nd suitble djustment, we hve the potentil to improve the verge system response time by 36.58% with the specific set of prmeters shown in Section 5.. Although burstiness is specil occurrence in network trffic, it sometimes hppens in rel situtions. According to our survey, most ppers set burstiness prmeter of rrivl rte (λ) or service rte (µ) to dd the burstiness fctor to M/M/ or G/M/ [4 6]. In the future we will refer to [4 6] or others to dd or modify some prmeters so s to consider the burstiness fctor in our model nd to verify it by simultion. References  Y.W. Bi & Y.C. Wu, Web dely nlysis nd reduction by use of lod blncing of disptcher-bsed web server cluster, IASTED Int. Conf. on Prllel nd Distributed Computer nd Networks, Innsbruck, Austri, 003,  V. Crdellini, M. Coljnni, & P.S. Yu, Dynmic lod blncing on web-server system, IEEE Internet Computing, 3 (3), 999,  M. Zri, H. Siedin, & M. Neem, Understnding nd reducing web delys, IEEE Computers, 34 (), 00,  H. Bryhni, A comprison of lod blncing techniques for sclble web servers, IEEE Networks, 4 (4), 000,  M. Cstro, M. Dwyer, & M. Rumsewicz, Lod blncing nd control for distributed World Wide Web servers, Proc. 999 IEEE Int. Conf. on Control Applictions,, Kohl Cost- Islnd of Hwii, USA, 999,  S. Ndimplli & S. Mjumdr, Techniques for chieving high performnce web servers, Proc. Int. Conf. on Prllel Processing, Toronto, 000,  L. Cherksov & S.R. Ponneknti, Optimizing contentwre lod blncing strtegy for shred web hosting service, IEEE Proc. 8th Int. Symp. on Modeling, Anlysis nd 88 Simultion of Computer nd Telecommuniction Systems, Sn Frncisco, CA, 000,  D. Mnjunth, D.M. Bhskr, H. Thilrmni, S.K. Bose, & M.N. Umesh, The queueing network nlysis tool (QNAT), Proc. 8th Int. Symp. on Modeling, Anlysis nd Simultion of Computer nd Telecommuniction Systems,, Sn Frncisco, 000,  H. Shin, S.H. Lee, & M.S. Prk, Multicst-bsed distributed LVS (MD-LVS) for improving sclbility nd vilbility, IEEE Proc. 8th Int. Conf. on Prllel nd Distributed Systems, KyongJu City, Kore, 00,  L. Cherksov, FLEX: Lod blncing nd mngement strtegy for sclble web hosting service, Proc. 5th IEEE Symp. on Computers nd Communictions, ISCC 000, Antibes-Jun Les Pins, Frnce, 000, 8 3.  L. Cooper & D. Steinberg, Introduction to Methods of Optimiztion (Phildelphi: W.B. Sunders, 970).  L. Avers & A. Bestvros, Lod blncing cluster of Web servers using distributed pcket rewriting, Proc. of the IEEE Int. Performnce, Computing, nd Communictions Conf., IPCCC 00, Phoenix, Arizon, 000, 4 9.   J. Abte, Asymptotic for stedy-stte til probbilities in structured Mrkov queuing models, Stochstic Models, 0, 000,  J. Rochol & M.H. Diemer, Adptive description of ATM trffic flows, CLEI Electronic Journl, 55 (), 000, 3.  Y. Cheng, X. Ling, L. Ci, W. Song, W. Zhung, X. Shen, & A.L. Grci, Sttisticl multiplexing, dmission region, nd contention window optimiztion in multiclss wireless LANs, Proc. 3rd Int. Conf. on Qulity of Service in Heterogeneous Wired/Wireless Networks (QShine) 006, Wterloo, Cnd, 006, Biogrphies Ying-Wen Bi is professor in the Deprtment of Electronic Engineering t Fu-Jen Ctholic University. His reserch focuses on mobile computing nd microcomputer system design. He obtined his M.Sc. nd Ph.D. degrees in electricl engineering from Columbi University, New York, in 99 nd 993 respectively. Between 993 nd 995 he worked t the Institute for Informtion Industry, Tiwn. Yi-Cho Wu is currently working towrd his Ph.D. t the Grdute Institute of Computer nd Communiction Engineering, Ntionl Tipei University of Technology, Tipei, Tiwn. He received his M.Sc. degree from the Deprtment of Electronic Engineering, Fu-Jen Ctholic University, in 003 nd his B.Sc. degree from the Deprtment of Computer Science nd Informtion Engineering, Chung Hu University, in 00. His reserch focuses on nlysis nd improvement of lod blncing on web servers, wireless d-hoc networknd wireless sensor network.