Nice Resource Reservations in Linux

Transcription

1 Nice Resouce Resevations in Linux Matin Ohlin Depatment of Automatic Contol Lund Univesity Box 118, SE-221 Lund, Sweden Matin A. Kjæ Depatment of Automatic Contol Lund Univesity Box 118, SE-221 Lund, Sweden Abstact Computing systems ae becoming moe and moe complex and poweful evey yea. It is nowadays not uncommon to un seveal seve applications on the same physical platfom. This gives ise to a need fo esouce esevation techniques, so that administatos may pioitize some applications, o customes, ove othes. This aticle gives a bief intoduction to the Linux kenel 2.6 task schedule. The aticle also pesents an implementation of a scheduling mechanism, that in a non intusive way intoduces CPU bandwidth esevations fo a task, o a goup of tasks, in the GNU/Linux opeating system. The scheduling mechanism is fist tested using dedicated load tasks,andthenonasetupconsistingoftwoapacheseves. I. INTRODUCTION Resouce esevation has become an impotant tool fo moden IT systems. Fo example, a Intenet host(e.g. a web hotel) guaantees to supply a cetain amount of esouces to a numbe of sevice povides(e.g. web shops). Often seveal sevice povides ae hosted on the same hadwae, but the host must guaantee that each sevice povide eceives the ageed amount of esouces, despite the behavio of the othe sevice povides.thespecifictypeofesoucescanbeoneomoe of; netwok bandwidth, database access, memoy allocation, CPU bandwidth, and many moe. Anothe example is when a movieplayeonapcneedsacetainamountofcpuesouces while a vius scanne uns in the backgound. On a single compute, the exact decision of how the esouces ae split between the applications is often left to the opeating system. In most cases, this defeence of command canbeseenasanadvantagebecauseitisnotnomallyknown exactly how impotant they ae elative to each othe. Fo example, it is not tivial to detemine how impotant the mail seve is compaed to the web seve. Howeve, in some cases, it would be advantageous if thee existed a mechanism to specify exactly how impotant diffeent tasks(o goups of tasks) ae compaed to each othe. Take fo example the simple casewheetwowebsevesaeunningonthesamemachine. Thefistseveisusedbypayingcustomes,andtheotheby bowsing customes. Then it would be advantageous to have a mechanism to contol the aveage esponse times, so that the paying customes will always be satisfied, wheeas the bowsing customes ae still allowed to access the seve if thee is spae capacity. The so called vitual hosting povides utilities to define esouce allocation by imposing a vitual opeating system between the host opeating system and each sevice povide. The esouces allocated to each vitual seve can be defined by the host. This gives well defined esouce isolation, but in some case, a simple, and possible cheape, method might be pefeed. Theesults descibedinthis papeaimto obtaincpu bandwidth sepaation between diffeent sevice povides while keeping the ovehead to a minimum. A feedback based method is used to achieve CPU bandwidth esevation on a kenel level, thus avoiding the need to make modifications to the applications. Because it is implemented in the kenel with an optimized algoithm, the ovehead is faily small. The implementation makes use of the Linux pioitizing scheme to assueaspecifiedamountofcpubandwidthtoagiventask. The CPU esevations ae obtained using existing opeating system infastuctue. AsthesoucecodeofGNU/Linuxisfee,anumbeof oganizations have eleased thei own vesions of the opeating system. Some of these oganizations ae Debian, Fedoa and SuSE. GNU/Linux has fo many yeas been a lage competito in sevesystems, suchas web,ftp,mailandfile seves. Duing the last yeas, the inteest in GNU/Linux fom commecial companies has inceased damatically. Nowadays, lage entepises such as IBM and Hewlett Packad take active pat in the GNU/Linux development, and also ship GNU/Linux as pat of thei lage seve and cluste systems. The emainde of this pape pape is oganized as follows. Section II descibes the objective of the contol mechanism. Section III descibes some issues of the Linux schedule elevant fo the esouce allocation poblem. In Section IV a model of the schedule is fomed, and Section V descibes the contolle implementation. In Section VI the contol algoithm is expandedto take the task s state into account to avoid integato windup. Section VII pesents expeimental validation using test tasks, and Section VIII pesents expeimental esults whee the esevation mechanism was implemented on a setup with two Apache seves. Related wok is pesented in Section IX, and finally, conclusions ae stated in Section X. II. OBJECTIVE OF CONTROL TheobjectiveofthecontolinthispapeistoachieveCPU bandwidth esevations. Such esevations make it possible to esevefactionsofthecputo specifictasksoagoups of tasks. In the Linux kenel, both pocesses and theads ae called tasks. Even POSIX theads ceated using the Native POSIX Thead Libay(NPTL) ae called tasks. Fom the

2 kenels pespective, they ae all schedulable entities. Theefoe,themethodwhichwillbepesentedcanbeusedbothfo applications made up of theads and of pocesses. The developed method has been implemented as an add on tothelinux 2.6kenel.Akeyfactointhecuentimplementationhasbeentomakeitnon intusiveandtopesevethe way that the oiginal schedule woks. This gives the benefit that the new featues can be used without compomising existing functionality. The CPU bandwidth contolle uses the as a contol signal, and the tasks execution time as a measuement signal. This foces the schedule to give the contolled tasks thei specified amount of CPU bandwidth. Itmaybeaguedthatthepesentedpoblemcanbesolved offlinebyspecifyingastaticnicevaluefoeachtask.thisis of couse absolutely tue, if the system is static and eveything isknownbefoehand.thatis,ifitisknownexactlyhowmany tasks that ae pesent in the system, and also thei execution timedemands.thesepemisesaenotlikelytoshowupin an odinay Linux desktop o seve system and theefoe it isnecessaytointoduceafeedbacklooptocopewiththe unknown.inanodinaycomputesystemtheeisalotof dynamics.thisisduetothefactthattaskscanaiveand leavethesystematanytime.taskscanalsochangetheistate andinthatwayconsumemoeolessexecutiontime.when unning on multi pocesso systems, tasks will also jump between pocessos in a(fom a spectatos pespective) moe o less andom patten. This causes the execution envionment to change apidly and theefoe an ability to adapt to diffeent situations is necessay. III. SCHEDULING OF TASKS Thee ae thee diffeent scheduling policies available in Linux; SCHED_FIFO, SCHED_RR, and SCHED_OTHER. The two fome ae fo soft eal time scheduling policies, and the latte is fo nomal time shaing scheduling. Only the latte is utilized in the wok pesented in this pape. The Linux 2.6 kenel task schedule has two pioity queues, onefoactivetasks,andonefoexpiedtasks.thequeues ae aays of linked lists, one list fo evey pioity. The schedule always chooses the list with the highest pioity intheactivequeue.eveytaskintheactivequeueofthe systemgetstounacetaintimebefoeitisputintheexpied queue.thistimeiscalledatime_sliceandtheactual sizeofitdependssolelyonthe nicevaluegiventothetask and not on the effective pioity. Nice values in the inteval [ ]aemappedtotimeslicesizesintheinteval [8 ms...1 ms... ms].thesizeoftheesultingtime slicescanbeseeninfigue1.notethattheesultingtime slices do not scale linealy with the. A non inteactive task is moved to the expied queue when ithasusedupitsgiventimeslice,aninteactivetaskonthe othehandiseinsetedintotheactivequeueiftheeisno iskofstavationintheexpiedqueue.whentheeaeno moetasksleftintheactivequeue,itisswitchedwiththe expied one. The expied queue is consideed to be staving ifthefistexpiedtaskhashadtowaitfomoethanafixed time_slice (ms) Fig.1. Thesizeof time_sliceasafunctionofthe nicevalue.notice that the s ae odeed fom negative to positive values to indicate inceasing time_slice allocation. time multiplied with the numbe of tasks in the active queue. This makes the stavation limit load dependent. It is also consideedtobestavationifataskwithlowe nicevalue than the cuently unning task is in the expied queue. Tasks with the same pioity ae teated diffeently, depending on whethe the schedule consides them to be inteactive o not. Non inteactive tasks ae not inteupted by tasks with the same pioity duing execution of thei time slice. Inteactivetasksontheothehandgettheitimeslicesplitup intosmallepieces,andaeputattheendoftheactivequeue again and again until they have executed fo thei whole time slice. The esult is that inteactive tasks ae scheduled moe o less ound obin with tasks of the same pioity, while non inteactive tasks un non peemptively. Fo a moe detailed desciption of the task schedule, see[1]. 1 IV. MODELOFTHESYSTEM Tostatwith,weassumethatataskalwaysneedstoun, i.e.,itiscpu bound.wecanthensummaizetheabovein afewpointsandbuildamodelofthesystemtopedictits behavio. Atask stimeslicedependssolelyonits nicevalue. Tasks ae scheduled in ode of pioity. Tasksaemovedfomtheactivequeuetotheexpied queue when they have executed thei whole time slice. The active queue is swapped with the expied when empty. Accoding to the model, the faction of execution time that ataskgetsduingoneoundofexecutioniscalculatedas: faction(i) = time slice(i) time slice(j) j whee iand jdenotetaskindices. 1 2

3 % Fig. 2. Compaison of the execution time model fo Linux (+) and measuements on a compute( ). Notice that the s ae odeed fom negative to positive values. Example:If,foexample,tasks 1, 2and 3havenicevalue andtask 4has nicevalue 1,thentask 4willget: f action(i) = A. Evaluation of the Model % To show that the model is accuate, theoetical values fom the poposed model ae compaed to values eceived though measuements. The expeimental setup consists of fou tasks unning in endless while loops. Thee of the tasks have nice value,whilethe nicevalueofthefouthtaskisvaiedin odetogiveitmoeolesscpubandwidthcompaedto theothetasks.theesultcanbeseeninfigue2.using lowe s than 6 esulted in the system becoming to sluggish to do any good measuements, theefoe these ae leftout.thissluggishnessispobablyduetothefactthat taskswiththatlow nicevaluesaegivensohighpioities that they conflict with the tasks inteacting with the use. As can be seen, the esults fom the expeiments follow the model vey well. V. CONTROLLER IMPLEMENTATION Thecontolsignal,inthiscasethe nicevalue,canonly change in discete steps. This makes it impossible to keep some efeences statically. But by using modulation techniques, as fo example Pulse Width Modulation(PWM), the efeence can be kept on aveage. Anothe thing that one shouldbeawaeofisthenonlineaityinhowthe nicevalue ismappedtothe time_slice,asseeninfigue1.this makes it hade to follow efeences which foces the contol signal to oscillate ove the inteval [, 1]. Ongoing wok is to compensate this by an invese nonlineaity, and theeby educing the oscillations. This will equie moe code to be executedinthekenel,buttheoveheadisexpectedtobe faily small. 2 A PI contolle with anti windup has been implemented using fixed point aithmetic in a kenel module. The PI contolles have a stong histoy in the contol community because it combines obustness and fast esponse, while being elative simple to configue, even without exact knowledge of the system to be contolled. In the futue, moe elaboate schemes could be used that take into account moe details of how the schedule woks, as well as global knowledge ofthebehavioofothetasks.byusingapi contollein this way, a PWM like behavio is achieved automatically. The implemented PI contolle is given by u(k) = K ( y ef y(k) ) + I(k) (1) K ( I(k) = I(k 1) + T S yef y(k) ) T i (2) whee the paametes K and T i ae the popotional and integal contol paametes, espectively. The vaiable u is the contol signal(the ), and y is the measuement (thefactionoftime giventothe task).thevaiable k is thediscetetimeindex,suchthat t = kt S,whee T S isthe sampling time. The PI contolle consists of two components. Thepopotionalpat(K(y ef y(k)))ensuesfasteaction to distubances, but does not assue that the desied efeence is eached. The integal pat(descibed by I) will accumulate any eo between the measuement and efeence in a simila manne as an incemental contolle. This pat is paticula beneficial when the system unde consideation is not well known and pedictable, as with compute systems. The specific implementation uses the contol paametes K =.1 and T i = 2,andsamplingtimesof T S = 2ms. Sincethecontoldesignisnotbasedonadynamicalmodel, thee ae no theoetical guaantees fo pefomance o stability. Howeve,thevaluesof K and T i havebeenchosenathe consevatively to have lage stability magins. This is imposed becauseevensmallchangestothe nicevaluecanleadto lage changes in the achieved CPU bandwidth. The effect of changestothenicevaluealsovaieswiththenumbeoftasks inthesystemandtheiespective nicevaluesintun.the fact that thee ae unknown paametes in the system makes it good to have a consevatively tuned contolle. Anothe thing that makes a consevatively tuned contolle pefeable, is the fact that measuements ae not taken instantaneously, but ove aninteval.thesystemmusttheefoebegivensometimeto eact to the contol signal befoe changing it again. Using a lagevalueon T i,alsohasthebonuseffectthatmeasuements ae aveaged. In the case whee the contolle satuates, the contol objective might not be met, since the contolle lacks actuation possibility. The integal pat will emain within the allowed contol ange due to the anti windup scheme. Fo moe details on the contolle implementation, see[1]. VI. TAKING THE TASK S STATE INTO ACCOUNT Up until now it has been assumed (fo simplicity) that a contolled task is always willing to un, i.e., it is cpu bound.thismaybetueinsomecases,butobviouslynot in all. Imagine fo example that the contolled task is given

4 aefeenceof %,butdoesnotneedmoethan 4%due tothefactthatitiswaitingonsomei/otooccutheestof thetime.theintegalpatofthepicontollewillthenadd up the diffeence, and incease the contol signal in ode to emovetheeo.butasthetaskisunwillingtoun,andthe systemcannotfoceit,theeowillemainandthecontol signalwill,duetotheintegaleffect,continuetoiseuntilit hitsitslimit.thisisofcousenotasatisfactoybehavio,and couldbeavoidedbytakingthecuentstateofthetaskinto account when contolling it. The stategy could be something like:donotinceasethecontolsignalfutheifthetaskisnot notwillingtounmoe.thisismoeolessananti windup scheme which ensues that the integal pat does not wind uptyingtoenfocehighecpuallocationtoataskthanthe taskdemands.theemainingpatofthissectionwillgivea stategy fo solving these kinds of situations. A. Stategy Theideatoupdatethecontolsignalonlyifthetaskis willingtoun,soundsgoodatfist.itis,howeve,notassimple asitfistmightseem.theobviousquestiontoansweis:how doweknowifataskiswillingtounmoethanitaleady does? The idea used in the cuent contolle implementation istosamplethestateofthetaskatthesametimeasthe execution time. The contolle is then only executed if two consecutive samples show that the task is in the unning state.thisstategywokswellifthetaskisusuallyinthe unning statefoalongetimethanthetimebetweentwo consecutive samples of the contolle. How long time a task spends in its unning state depends highly on its wokload duingthattimeinteval,butitalsodependsontheothetasks inthesystem,asthetaskmightgetinteuptedbyahighe pioitytask.thismakesithadtogiveanygenealules andhencedawanyconclusionstobeusedfomoeaccuate contol. B.WhydoestheStategyWok? The eason why the stategy woks, i.e., the contol signal doesnotsatuate,isthefollowing:ataskwithalowpioity willbeinthe unning statefoalongtime.thisisdueto thefactthatitwillbepecededandinteuptedbytaskswith highe pioities. It will not switch fom the unning state untilithasfinisheditscuentwokload.ifthepioityofthe taskisinceased,thetaskmaynotbepecededbyasmany tasksasbefoeanditwillalsonotbeinteuptedbyasmany. Hence,itwillfinishealieandtheefoebeashotetimein the unning state. In essence, a high pioity gives a shot timeinthe unning state.astheexecutiontimedemandis constant, the atio between executed time and time spent in the unning state will incease if the pioity is inceased. At a cetain point, an equilibium will be eached, whee the efeenceismetduingthepeiodwhenthetaskisinthe unning state, and hence the contol signal will be constant. Example: Fig. 3 shows the sampling points of the contolle, andthetask sstateatthosepoints.italsoshowsthetask s execution tace and which of the sample intevals that ae used Task Execution Sampling Points Task State Contolled Intevals CPU Bandwidth w % % % 7% 1% 3% 9% 3% % Fig. 3. Figue showing how the stategy descibed in Section VI-B fo contolling non cpu bound tasks woks. bythecontolle.an meansthatthetaskisinthe unning state,andawthatitisinoneofthe waiting states.duing the contolled intevals, the atio between executed time and time spent in the unning state is appoximately 48%. VII. EXPERIMENTS WITH LOAD TASKS Expeiment1and2havebeenpefomedonasingleCPU desktop compute. At the same time as the expeiments wee made, thee wee a numbe of tasks in the system, e.g., X, Fiefox, Thundebid, XEmacs and so on. All bandwidth measuements have been filteed though a moving aveage windowof 4s.Thefilteingisdonebecauseofthefactthat whenataskexecutes,itgets 1%oftheCPU bandwidthand thenitgets %whenitdoesnotexecute.filteingthougha moving aveage window shows the CPU bandwidth duing thatwindowandthisisalsowhatonewantstoachieve. Running the expeiments on a compute with moe than one CPUwillgainesultssimilatotheonesseeninthissection, except that thee will consideably moe load distubances, as those seen in Expeiment 1. Expeiment 1: The setup in this expeiment consists of fou tasks unning in endless while loops. Two of the tasks have theinicevaluessettofive,andactasbackgoundload.the thid and fouth task s nices values ae used as contol signals to keep the measued bandwidth at the desied efeences. Theefeencesfobothofthetasksaekeptat 2%initially. Attime 182,theefeencefothefisttaskischangedfom 2%to %.Ataoundtime 32,theefeenceischanged backto 2%.Theesultofthestepesponsefothefisttask canbeseeninfigue4.thecouplingbetweenthetwotasks isvisibleinfigue,whichshowsthedistubanceonthe secondtaskesultingfomthesteponthefistone.nofeed fowad tem is used in the contolle. This expeiment shows thepwmnatueofthecontolsignalandtheesultsofthe quantizationinthe nicevalue.infigue4,itcanbeseen thatthecontolsignalisconstantbothbefoeandaftethetwo steps.butwhentheefeenceissetto %,thecontolsignal fluctuates a lot. Also note that thee is much less oscillation inthecpu bandwidthwhentheefeenceissetto 2%than to %.Thisisduetothefactthatsomeefeencescannotbe kept stationay because the is discete. Expeiment 2: This expeiment consists of one peiodic task that executes fo appoximately 4 ms and then sleeps fo 6msepeatedly.Thisesultsinataskthatusesatmost 4% ofthecpuevenifitisaloneinthesystem.contollingsuch ataskequiesthestateofthetasktobetakenintoaccount asdescibedinsectionvi.twoloadtasksofthesametype usedinexpeiment2aealsopesentinthesystem. w w

5 7 6 CPU Bandwidth of Task 1 Measuement Refeence 8 6 CPU Bandwidth Measuement Refeence % 4 3 % Contol Signal of Task Contol Signal Fig. 4. Step esponse of the CPU bandwidth(task 1) when contolling two tasks in Expeiment 1. % CPU Bandwidth of Task Contol Signal of Task 2 Measuement Refeence Fig.. Step esponse of the CPU bandwidth(task 2) when contolling two tasks in Expeiment 1. AscanbeseeninFigue6,thepoposedschemewoks wellinpactise.inthebeginningoftheplot,theefeenceis highethanthetaskdemands,andattime itissettoan even highe value, but the contol signal still behaves well. Itcanalsobeseenthatthecontolleisstillabletofollow efeence changes when they ae lowe than 4%. The obsevant eade may notice the delay and the following unde shoot at time 16. Also note that this behavio does not showupatanyoftheothestepchangesintheplot.this isnotanintegatowindupasmightfistbethought,butis insteadduetothefactthatthesystemhasmakedthetaskas inteactive and theefoe given it an additional bonus. When the task afte some time is maked as non inteactive, it loses its bonus and this esults in the unde shoot Fig.6. Stepesponsesfoanoncpu boundtaskwhentakingthetask sstate into account in Expeiment 2. VIII. EXPERIMENT WITH APACHE SERVERS This expeiment aimed to test the scheduling mechanism on a moe ealistic application. The test is not included to suggest that the mechanism is the best solution fo the given example, butonlytodemonstatethatthemethodwoksfoapoblem including moe advanced behavio than the pevious tests. The setup epesents a hosting system whee two sevice povides ae hosted on the same physical hadwae. The objective is to sepaate the two sevice povides such that one sevice povide can opeate unaffected of a equest oveload at the othe sevice povide. Expeiment3:APentium4,1GBmemoy,3GHzPC, with a Linux Fedoa opeating system and kenel was used as host compute. Two Apache seves, vesion 2.2.2, configued by using the pefok module, wee installed on the seve as two distinct sevice povides. Using this configuation, a equest was handled by one pocess(child). If all existing pocesses wee occupied, Apache dynamically stated new pocesses, and likewise, closed some if thee wee too many idle pocesses. This means that the numbe of pocesses associated with one Apache seve changed ove time. The contolle was configued to set a common nice valuetoallthepocessesofagivenapacheseve.also,the timeallocatedtoallthepocessesofoneapachesevewee summed to give the time faction measuement. In this manne, a single input single output system was obtained as equied by the contol stuctue. Only one of the Apache seves was contolled with the poposed scheduling mechanism. The contolled and uncontolled seve wee listening on pot 8 and 81, espectively. The setup is illustated in Figue 7. Taffic was geneated fom 12 client computes(athlon, 1. GHz PC), gouped into thee equal goups consisting of fou client computes each. The taffic was geneated using the taffic geneation softwae CRIS[2]. All clients equested the same PHP file(geneating a esponse with 7 chaactes)

6 Aival intensity Host compute Aival intensity Aival ate Time Client #1 Contolled Seve Pot 8 Uncontolled Pot 81 TCP/IP laye Seve 1 Mbit switched netwok Client #12 Fig. 7. Expeimental setup fo expeiment 3. with exponentially distibuted inte aival times, and wee configued to timeout afte 1 s. Duing the expeiment, all the computes wee connected on a 1 Mbit switched Ethenet netwok. At the beginning of the expeiments, client goup I stated sending equests to the contolled seve and client goup II stated to sent equests to the uncontolled seve. Both goups sent appoximately 16 equests/s. This taffic did not give ise to CPU oveload, but left appoximately 2% CPU bandwidth fee. A seve is consideed to be oveloaded when the equests cannotbesevedwithinthetimeoutoftheclientsduetolack ofcpuesouces.afte173s,clientgoupiiistatedtosend appoximately 16 equests/s to the uncontolled seve. The compute did not have sufficient CPU capacity to maintain opeation of both seves. The aveaged aival intensity is shown in the uppemost sub figue of Figue 8. The taffic going to the uncontolled seve was the combined taffic fom clientgoupiiandiii. Unde the initial opeation none of the seves wee oveloaded but as the equest ate to the uncontolled seve was doubled, the compute lacked the CPU bandwidth to seve all equests. Pefeably, only the seve being exposed to the exta taffic should become oveloaded, while the othe seve should emain opeational. Themiddleandbottomsub figuesoffigue8showthe esponse times of the contolled seve and the uncontolled seve, espectively. In the case whee the contolle was inactive, both seves became oveloaded when the taffic inceased. Afte the incease of taffic, the esponse times of both seves inceased damatically and all clients stated to timeout. Consistent timeouts fom the clients wee obseved. In the case whee the CPU esouce allocation was contolled by the poposed scheduling mechanism(efeence set to 4% CPU bandwidth), only the uncontolled seve became oveloaded. The contolled seve continued to pefom with simila esponse times. Client timeouts wee obseved consistently only on the uncontolled seve s equests. Two single timeouts wee obseved on the contolled seve s equests. The small Time Aival ate (eq/s) 3 2 Contolled seve Uncontolled seve Time (s) Response time fo contolled seve 6 Inactive contolle 4 Active contolle Response time (ms) Response time (ms) Time (s) Response time fo uncontolled seve 6 Inactive contolle 4 Active contolle Time (s) Fig. 8. Results fom expeiment 3 on two Apache seves. Top: Aveage aival ates. Middle: Response time fo the contolled seve with and without feedback contol. Bottom: Response time fo the uncontolled seve. All vaiables wee measued at the clients and wee filteed with a moving aveage window of 2 equests. jump obseved in the esponse time of the contolled seve (middlesup figueoffigue8)ataound2sisassumedto beduetosomedistubancefomothetasksintheopeating system. This expeiment showed that the scheduling mechanism can be used to affect the esponse time of a web seve application. The setup with two diffeent seves with diffeent client goups can be used in applications whee diffeent sites ae hosted on the same physical compute, but whee the pefomance of one seve must be independent of the behavio of the othe seve. In this expeiment we have not consideed howsuchasystemshouldbesetupinaealapplication.fo instance,wedonotconsidehowthetaffictothetwoseves is sepaated. Howeve, the expeiment shows that the two seves can be sepaated by means of the poposed scheduling mechanism. The setup does not aim to contol the esponse time.ifthiswastheobjective,asecondcontolloopwould have to be included, defining the efeence fo the CPU bandwidth contolle. IX. RELATED WORK Resevation based scheduling is not a new concept and has been aound in one fom o anothe fo many yeas. The concept has been called fai shae scheduling[3],[4], [] but is also known unde the name popotional shae scheduling[6],[7],[8],[9].agoodsummayofthisfield, togethewithmoedetailscanbefoundin[1]. TheideaofusingthenicevalueasawaytoenfoceCPU factions has been known befoe. One ealy implementation is the Watson Shae schedule [11], implemented on top

7 of a standad AIX opeating system at the Compute Powe SeveClusteatIBM.Itisalsomentionedin[12]and[13] as something that in UNIX can be done in theoy,but is complicated in pactice because of the non linea elationship between nice, the numbe of pocesses and the CPU faction eceived.povidedthatthenumbeofjobsinthesystemis fixed,andthattheyaeallpesentfomthesametimeand onwad, a deteministic analysis of the steady state shaes is possible.[14] shows how this can be used to statically calculate the base pioities on a unipocesso in the pesence of decay usage scheduling in UNIX.[1] extends this analysis to the multipocesso case. An inteesting Linux kenel poject in this aea is Class based Kenel Resouce Management(CKRM)[16] and[17] which aims at poviding diffeentiated sevice to esouces such as CPU bandwidth, memoy pages, I/O and incoming netwok bandwidth. It accomplishes CPU esevations by scaling the time_slice value and e queuing tasks. Pats ofthispojectisusedin SuSELinuxEntepiseSeve9, but not the CPU contolle. Resouce allocation and esevation in web seve applications is nothing new. An example is[18] whee vitual seving on one physical compute is used to guaantee cetain quality of sevice metics fo seveal client classes unde chancing load conditions. An often used actuation method is admission contol, whee equests ae denied in ode to avoid oveload and to guaantee cetain pefomance metics fo the accepted equests[19],[2],[21]. X. CONCLUSIONS AnexpositionoftheLinux 2.6schedulehasbeendoneand a feedback based method fo contolling the CPU bandwidth given to tasks in Linux has been pesented. The pesented methodhasbeenshowntowok,bothfocpu boundand non cpu bound tasks. A numbe of expeiments have been pefomedinodetoshowthatthetechniquewoksineality. The expeiments indicate that CPU bandwidth allocation can be obtained with the poposed scheduling mechanism. Futhemoe, the mechanism can be used to sepaate the pefomance of two Apache seves unning on the same physical compute that is, one seve emains opeational while the othe seve is oveloaded. [7] I. Stoica and H. Abdel-Wahab, Ealiest Eligible Vitual Deadline Fist : A Flexible and Accuate Mechanism fo Popotional Shae Resouce Allocation, Nofolk, VA, USA, Tech. Rep., 199. [8] C. A. Waldspuge and W. E. Weihl, Stide Scheduling: Deteministic Popotional-Shae Resouce Mangement, Massachusetts Institute of Technology, MIT Laboatoy fo Compute Science, Tech. Rep. MIT/LCS/TM-28, June 199. [9], Lottey Scheduling: Flexible Popotional-Shae Resouce Management, in Fist Symp. on Opeating Systems Design and Implementation(OSDI). USENIX Association, 199, pp [1] J. de Jongh, Shae Scheduling in Distibuted Systems, Ph.D. dissetation, Delft Univesity of Technology, 22.[Online]. Available: d/dejongh.pdf [11] C.MouzziandG.Rose, WatsonShaeSchedule, inpoc.ofthe Fifth Lage Installation Systems Administation Conf.(LISA 91). San Diego, USA: USENIX, 1991, pp [12] J. L. Hellestein, Challenges in Contol Engineeing of Computing Systems, in Poc. of the 24 Ameican Contol Conf., vol. 3, 24, pp [13] J. L. Hellestein, Y. Diao, S. Paekh, and D. M. Tilbuy, Contol Engineeing fo Computing Systems, IEEE Contol Systems Magazine, vol.2,no.6,pp.6 68,dec2. [14] J. L. Hellestein, Achieving Sevice Rate Objectives with Decay Usage Scheduling, IEEE Tans. Softwae Eng., vol. 19, no. 8, pp , [1] D. H. J. Epema, Decay-Usage Scheduling in Multipocessos, ACM Tans. Comput. Syst., vol. 16, no. 4, pp , [16] CKRM, Class-based Kenel Resouce Management(CKRM), Home page: [17] S. Naga, R. V. Riel, H. Fanke, C. Seethaaman, V. Kashyap, and H. Zheng, Impoving Linux esouce contol using CKRM, in Poc. of the 24 Linux Symp., vol. 2, Ottawa, Ontaio, Canada, July 24, pp [18] W. Xy, X. Zhu, S. Singhal, and Z. Wand, Pedictive contol fo dynamic esouce allocation in entepise data centes, in Poc. Conf. on Management of Integated End-to-end Communications and Sevices, NOMS, Vancouve, Canada, Apil 26, pp [19] X. Chen, H. Chen, and P. Mohapata, Aces: An efficient admission contol scheme fo qos-awae web seves, Compute Communications, vol. 23, no. 14, pp , 23. [2] S.C.Lee,J.C.Lui,andD.K.Yau, Apopotional-delay diffsevenabled web seve: admission contol and dynamic adaptation, Paallel and Distibuted Systems, IEEE Tans., vol. 1, no., pp. 38 4, 24. [21] M. Andesson, J. Cao, M. Kihl, and C. Nybeg, Admission contol withsevicelevelageementsfoawebseve, inpoc.ofiastedint. Conf. on Intenet and Multimedia Systems and Applications(EuoIMSA), Gindelwald, Switzeland, Febuay 2. REFERENCES [1] M. Ohlin, Feedback Linux scheduling and a simulation tool fo wieless contol, Depatment of Automatic Contol, Lund Univesity, Sweden, Licentiate Thesis ISRN LUTFD2/TFRT SE, June 26. [2] M. Andesson, A. Hagsten, and F. Neisle, Cisis equest geneato fo intenet seves, in Poc. Fouth Swedish National Compute Netwoking Wokshop, Lule, Sweden, 26. [3] R.B.Essick, AnEvent-BasedFaiShaeSchedule, inpoc.ofthe Winte 199 USENIX Conf. USENIX, 199, pp [4] J. Kay and P. Laude, A fai shae schedule, Communications of the ACM,vol.31,no.1,pp.44,1988. [] G. J. Heny, The Fai Shae Schedule, AT&T Bell Laboatoies Technical Jounal, vol. 63, no. 8, pp , Octobe [6] L. L. Fong and M. S. Squillante, Time-Function Scheduling: A Geneal Appoach to Contollable Resouce Management, IBM Reseach Division, T.J. Watson Reseach Cente, Yoktown Heights, NY 198, Tech. Rep. RC 21(89194), August 199.