服务Web服务器基于不精确响应的Web服务机制_英文_

An Imprecise Response Service Mechanism

Based on Differentiated Service for Web Server

Hi UAfNlG Guii-mihn, lZHOUiliYa

(Guilin Univers ty o E ectron c Tec no ogy,Gu n 541004,China)

requests into diferent level queues.Then, by means of an imprecise scheduler and a content adaptor, the

WQS could make requests in high priority queues to have good response rate and short response delay andcould make common requests in low priority queues to have satisfactory response rate and acceptableresponse delay. Finally, a series of simulating experiments for the WQS in Apache Web server wasdescribed.The experiments show that the WQS is a feasible and effective method to provide differentiatedservice and guaranteed quality of service for improving traditional Web server or Web services.

Keyl-twi ords:hWdebli qualitlyi of service;differentiated Web service; imprecise computation method;weakly-hardrea mesc e u ngpo cy

Web服务器基于不精确响应的Web服务机制

黄桂敏,周 娅

桂林电子科技大学广西桂林541004

摘 要传统Web服务器对于到达的请求使用先来先服务机制。但是 先来先服务机制无法为到达的请求提供区分

服务和服务质量保证。在不精确计算和弱-硬实时调度策略基础上提出了一种Web服务质量机制简称一WQ S为

到达Web服务器的请求提供区分服务与服务质量保证。WQS不同于先来先服务机制在于它首先利用 个请求分

类器分类到达的请求到不同级别的队列。其次借助不精确调度器与内容自适应器WQ S能够使优先级高的队列中

请求具有好的响应率和短的响应延迟使优先级低的队列中请求具有满意的响应率和可接受的响应延迟。最后本

论文描述了一系列在Apache Web服务器上针对WQS的仿真实验实验表明WQS是一种能够提供区分服务与服务

质量保证达到改进传统Web服务器或Web服务的可行和有效的方法。

关键词Web服务质量区分We b服务不精确计算方法弱-硬实时调度策略

中图分类号 TP39302 文献标识码 A 文章编号 1004-731X(2008) 18-4881-08

Introduction

With the widespread application of Web server, become a rallying cry for all Web servers on WorldWeb Quality of Service (Web QoS) is becoming an Wide Web application.However, the traditional Webimportant factor in distinguishing the success of Web server just uses first-co’me first-served (FCFS)service.Web QoS determines Web service usability mechanism to handle client s requests ’and FCFS is aand utility in Web server and it influences the popularity kind of fair mechanism for every client s req’uest.Thatof Web server on World Wide Web application. is to say, FCFS cannot differentiate client s requestsHowever,Web server is making an increasingly critical into different service classes according to their prioritiesissue because of unresolved critical Web service problem when these requests are being handled in commonto suffer from unacceptable levels of performance Web server. So traditional Web server not able todegradation.Actually,Web server is finding that is provide differentiate’d Web service and guaranteedWeb QoS for client s requests.For example,Apache

作务者与简P2介P计黄算桂周敏娅(1961956),6男女,广西湖,北博士硕,士副教授,研究方向为Web服 and all requests correctly received ar’e eventually

务与数据库技

第20卷第18期 Vol 20No 182008年9月 系 统 仿 真 学 报 Sep,2008

Apache Web server. In Apache Web server, these mechanism called Web2K server [5].And also, someprioritized requests cannot be protected from overload commercial products have been des’igned by ITWeb server.And implementing QoS in Web server is coimddplanies, sudchIBMas’ HWewbleStt-hPackaIrd sthi Web QoSa critical and significant challenge because of Web mi ewi are an t ti s e th pd ered. n s pkalperh, and

tmethokd anlid wt’eakly-hatrdt realt-timh e tschedulingti polifcy series of simulating experimen,ts for WQS based ono ma e c en s reques s o ge s or response me or Apache Web server.The simulating experiments showimproving FCFS to implement differentiated Web that WQS is a feasible and effective method toservice and guaranteed Web QoS in Web server. implement differentiated Web service and guaranteed

In the past years, in order to implement differentiated Web QoS in Web server on World Wide WebWeb service and guaranteed Web QoS for Web server, application.

can provide N classes of proportional-delay into two logical parts, a m,andatory part and andifferentiated services (PDDS) to heterogeneous clients. optional part. The mandatory part includes allWitfh PDDS,Weib server cani provide cfonsistent operations needed to produce logically correct results.

configurable fixed percentages ’of bandwidth across computation method takes to solve transient overloadnumerous simultaneous client s requests and ’is is to prematurely terminate or discard entire optional

prototype with priority based QoS control mechanisms but it should be acceptable if all the mandatory part iscalled QoS Apache. The QoS Apache server can completed. In order to ensure this, there must be aprovide different type of requests with differentiated rigid precedence constraint between these two parts.

approach that uses approximation techniques to execution.

第20卷第18期 Vol 20No 182008年9月 黄桂敏,等Web服务器基于不精确响应的Web服务机制 Sep,2008

Web server can execute differentia’ted Web service client’s requests under the situation of transient’according to the priority of client s requests or not. overload.And WQS can satisfy the attending client sBased on this idea, this paper designs a WQS requests to have the perception although with amechanism for common Web server.The WQS work reduced quality.That is to say, as soon as W’QS selectsidea as follow: first of all,WQS uses a request the best quality Web pages to respond client s requests,classifier to classify the incoming requests int’o different a precise execution has taken place in Web server.service classes that include different class s queues. Otherwise,when WQS sele’cts the degraded qualityThat’is to say, each service class is corresponded to a Web pages to respond client s requests, it can also beclass s queue.Then,WQS uses an imprecise scheduler said an imprecise execution to be occurred.Figure 1 isand a content adaptor,which are designed by imprecise the structure of WQS between operating system andcomputation method and weakly-hard real-time Web server software. In the context, it introduces thescheduling policy, to give a partial response for structure of WQS and implementation of WQS.

2 Imprecise Scheduler

Guillem Bernatuse proposed the notion of wi heththerkt-hte dleadAlilne isthmisslefdt(bitt=b0it)or met (btit=th1)weakly-hard real-time scheduling policy to represent nld te E uhp e. iso, eti e mos rheipftresfen lsl thean appropriate conceptual framework for specifying o es . ac new nvoca on causes a s o a ereal-time tasks that can tolerate occasional losses of bits towards left, the leftmost exits from the word anddeadlines.Tasks that must meet all of their deadlines is no longer considered,while the rightmost will be aare a particular case of their definition and will be 1 if the a request has met its deadline’or a 0 otherwise.referred to as strongly har[d10]real-time tasks whenever Definition 2:Let a class queue s µ-pattern isαthe distinction is relevant . In imprecise scheduler, and|α|≥m where m≤k. If aµ-patternα,with|α|≥m

From definition 1, it can be see thatµ-pattern is a ,4)Letβ=λ first leftλshif,tαthen input 1 from rightofα, thus CF (β)<CF (α),where|α|=m and|β|=m.

•4883 •

第20卷第18期 Vol 20No 182008年9月 系 统 仿 真 学 报 Sep,2008

In definition 3, the critical function CF λ(α) is mechanism with priorities higher than any other classpositive or zero if the constraint is satisfied and it will queues scheduled under the normal mode scheduler.still be satisfied in the future if all next deadlines are No restriction is placed on the normal mode scheduler.met. It is negative if it is not satisfied or, even 2)’Whenever a class queue is invoked, the classassuming that all future deadlines are met, if the queue s critical function is computed. If criticalconstraint will not be satisfied.The last two properties function is zero then the class queue may be in paniccapture the fact when the next deadline is met, the mode.Otherwise, it is in normal mode.critical function is more than before.On the contrary, 3) The transition from normal mode to panicif the deadline is missed, the criticalλfunction either mode is simply implemented as a priority promotion.decreases or remains the same. If CF (α)=0 for some 4)Whenever a request’ has been responded in aµ-patternα, then it can be said that the task is in a class queue, its class queue s pastµ-pattern is updatedcritical state. In this case, the following task accordingly.invocations have to meet their deadlines to satisfy the 5) If there are no class queues in panic mode thenwf eilakly-thard ictoi nsltrfaintt.iWhien thetitask has a dynamic the scheduler may choose a ready class queue to

is no risk to miss a weakly hard constraint.Obviously, d Iitb is hnecebssalry to uslie mafthemaiticail formhulal tonon-critical tasks will not produce a change to panic escr f t ei a lso ute qiua tyf oWseSrv cFei n efac llc ahssmode of imprecise scheduling policy. Because they qbueule or mlip emfentat oi n o f l Q . rst oi a ,kt deare not close to ge’nerate a dynamic failure. So the a sAobute qSua tyi bol serdv icte o l c aiss queules sbmarIfeimprecise scheduler s principle can be summarized in als sQo var ia t ei an is va lue s a reda knutm der.f n

panic mode scheduler and a normal mode schedul,er. queue can be calculated by the following formula(I):The panic mode is implemented as a fixed priority

•4884•

第20卷第18期 Vol 20No 182008年9月 黄桂敏,等Web服务器基于不精确响应的Web服务机制 Sep,2008

mostrightk=1,2, ,n If ProQoS(k) ≥ProSet(k)

slot of a class queue. HistoryLog(k, j+1) mostright Priority(k)=100-ProQoS(k)

Queue to its position and stored structure in Web server. In

k=1,2, ,n content adaptor to determine which tree to serve an

class queues, then the priority value of the kth class requests originated from the same client must be

•4885•

第20卷第18期 Vol 20No 182008年9月 系 统 仿 真 学 报 Sep,2008

served from the same content tree whenever possible. E}l

For the purpose, the hashing function maps a given se//No deadline

HTTP command library rather than via the TCP socket Else=library. The main pseudocodes of WQS can be /*GPoin1t;tresummarized in the following: Call HTTP ecoGm*m/an

/*The ith class queue has reached on critical status */ person’al computers.These thousands upon thousandsIf(Classi->AbsQoS<Classi->SetQoS client s requests arrive at listening queues in ApacheoCrlaCslsai-s>siS-t>aPturosQ=“oFS”<;/C/lCasrsiti-ic>aPlrsotSaetut)sThen Web server at through a local area network.Apache

Classi->AbsQoS=Classi->AbsQoS–Classi->HistoryLogoldest there are five service classes or class queues that is A,

+Classi->HistoryLoglatest; “ ” B,C,D,E and F. The load increment is 50request//C*alCl A subprogram to bhandlel class queue*s/tatus with i ; second to each class queue.The following figure 4 isLossoRuentuaesatm=oLunotssoRfe euinegst+os1t request some curves of simulating data from a series ofDelete tqhe ARequest froqm the Cl;assi->ClassQueue; simulating experiments. The simulating experiments

.

第20卷第18期 Vol 20No 182008年9月 黄桂敏,等Web服务器基于不精确响应的Web服务机制 Sep,2008

第20卷第18期 Vol 20No 182008年9月 系 统 仿 真 学 报 Sep,2008

33%(precise)and 36%(imprecise); the F class request QoSi middlewaretford iWmpblemeSniting differeWntiabted Webresponse rate is 18%(precise)and 22%(imprecise). serv ce an guaran ee e Qo n common e server.

On the whole, the results of simulating experiments References:shdowithat WthQS mani ages to redutice deadlinde misisets iby [1] Sam C M Lee, John C S Lui,David K Y Yau A Proportional-Delayre uc ng e serv ce mean mes an ma n a ns DiffServ-Enabled Web Server: Admission Control and Dynamicdifferentiated Web services and guaranteed Web QoS AS dsatpetmatsio(nS104[J5]-92I1E9E)E20T0r4an1s5a(c5ti)o:n3s85-o4n00Parallel and Distributed

of client’s requests have happened in Web server.On Ondi PerdforSmanc,e Evaliuatioin,of Ca /Q/oISE-EbaEseId Web iServl ers Skerhvice-the contrast,WQS always trie’s to maintain proportion 20r-e2n1teOct 2y0st0e5mUSnAg:nIEeeErEng200[5]: 191-196 nternat ona Wor s op,

5 Conclusion:

service and guaranteed Web QoS for Web server.On [7] J82Y(NCohu1n),gJaJnWuaryS 1L9iu94KUJSALi:nIESEcEh,ed1u9l9i4ng P

imprecise form and to distribute the deadline missed 9 DavceidmbMero2s0b0e1r,eLrondhotnt eUrfS-A: IETEooEl,2f0o0r5:M25e-a3s8urin

FCFS.Based on the paper, our further research work E0n0g0ineLeirminegr,icPkr,oIcreleadnidng, (sISoBf Nth:e12-50801013In-t2e0rn6a-9ti)o,n2a0l0C0:o2n6e3r-e2n7c2e, Junwill improve WQS and try to developa kind of Web

[王M小]英San, F赵ra海ncis陈co英:M革orga等n K异au构fm计an算n P系ub统lis任he务rs调In度, 1的99遗8传算法及 [10] AHleloLca,tiJoanrvisin S,MSuplotiocnluersteDrs,N[uCd]d/GProOcpeteimdiinsginsg oSftati1c8tWh orkIEloEaEd

[J] 系统,仿真学,报,2005, 17(4):798-800 Tnetcehrnaiqtiuoensaal ndCAonpfpelriecnactieonosn20P0a2rallel and Distributed Processing

•4888 •