A report on the performance benefits of upgrading memory on servers running VMware ESX Server 4.0

Transcription

1 2 h t cnario 3 t Procdur A rport on th prformanc bnfit of upgrading mmory on rvr running Vwar X rvr 4. 5 t Rult rvr Rpon im 6 Wb rvr mory Uag 7 ataba rvr mory Uag 8 rvr Faild ranaction 9 Concluion ting conductd and rport compild by Binary ting Ltd whavn ntrpri Cntr nton Iland whavn at ux B9 9BA t +44 () info@binarytting.com

2 t cnario hi rport wa commiiond by Kington chnology to look at th prformanc bnfit of upgrading from 2GB to 32GB of phyical mmory on rvr running Vwar X rvr and prnting multipl virtual machin (V) all configurd to function a wb and databa rvr. A uit of tt wa run to dmontrat th bnfit of mmory upgrad for th following cnario: 1) h amount of wb traffic multipl V can handl bfor prformanc dgrad. h tt wr conductd uing a pirnt tcntr applianc configurd to run multipl Avalanch. h gnrat ur rqut for wb contnt from th V and maur th rpon tim. 2) h prformanc of multipl V running databa rvr. h pirnt tcntr wa alo ud for thi tt whr imulatd ur rqut accd a backnd QL rvr databa running on th am V a th wb rvr. A rqutd by Kington chnology, th tt uit wr run for ach cnario whr th rvr undr tt wa fittd with 2GB of R-2 mmory and thn upgradd to 32GB compriing 4 x 8GB R-2 modul. RVR HARWAR PCIFICAI Chai: uprmicro 1U Rack ount mory: 667Hz R2 (2GB and 32GB) RAI: Intl B2 3Gbp AA CPU: 2 x 2.5GHz Intl Xon L542 quad-cor torag: 2 x 1B W driv in mirrord array twork: Intl 1GBaX 1Gb PCI- h rvr wa configurd with Vwar X rvr v4. and ach V configurd with icrooft Window rvr 28 ntrpri x64 running II 7.. For th backnd databa tt ach V wa alo configurd with icrooft QL rvr 28 x64. o avoid any ntwork bottlnck th rvr undr tt wa connctd via a fibr-optic 1-Gigabit uplink to a tgar G7328 thrnt witch. h pirnt tcntr had all ight port connctd to th witch coppr Gigabit port. 2

3 t Procdur VWar X rvr allow multipl virtual rvr to run on on phyical rvr. Although ach virtual machin (V) can b configurd mor or l a th ur rquir, thr ar om thing that can b don to tun X prformanc to obtain bttr rult. Although tuning X rvr for optimum prformanc i outid th rmit of th tt it wa poibl to do om impl tuning to mak bttr u of rourc. Idally, th hot rvr will hav nough main mmory to mt th total mmory rquirmnt of all th virtual machin a wll a it own mmory ovrhad. For 1 V ach configurd to u 2GB of mmory th hot ytm would nd mor than 2GB of RA. Howvr inc mot ytm do not u all th mmory availabl to thm, thi would man that th hot ytm RA would b undr ud. Whn a ral ytm i running a particular application it working t of mmory i gnrally l than th mmory availabl. A V thrfor only nd to b configurd with nough mmory for thi working t, which in turn rduc th amount of hot mmory ndd to run it. W took advantag of thi fact whn w t up th V and configurd ach on to u 1GB of mmory. W alo intalld th VWar ool o that th ytm could tak advantag of X rvr own optimiing function. X rvr ha a numbr of tchniqu to mak bttr u of hot mmory. h firt tchniqu i pag haring, which man that whn a V tart to load a pic of cod or data into it mmory X rvr chck to if it alrady ha a copy of that pag in mmory from anothr V. If it do hav a copy thn intad of allocating ral mmory to that pag it imply har th on it alrady ha. Whr vral V ar running copi of th am oprating ytm, thi aving can b ubtantial. h cond tchniqu i calld ballooning. Ballooning can only tak plac if th VWar ool ar intalld on th V. It allow th gut oprating ytm to communicat with th hot in uch a way that X rvr i awar of th mmory pag th gut oprating ytm ha frd up. X rvr can thn rla th ral mmory that it ha providd for th pag and rallocat thm a ndd. Without thi tchniqu mmory rmain allocatd to a virtual machin vn though it i not bing ud. h final tchniqu i to wap pag to dik. X rvr u thi whn th hot mmory i ovr-committd. Although om wapping may occur whn tarting and topping virtual machin, wapping whil ytm ar running may indicat a nd for mor hot mmory. xtniv wapping will alo hav an impact on ovrall prformanc, which will vntually b rflctd in poorr rpon tim. trmining whn prformanc ha dgradd to unaccptabl lvl i to om xtnt a ubjctiv proc, but w cho om ky indicator o that w could dfin whn thi point had bn rachd for th tt. Indicator includ high rat of tranaction failur for whatvr raon (tranport rror, tim out, ur abort), high lvl of wapping in X rvr, xciv rpon tim, failur to achiv th pcifid workload. Whn w ttd th wb rvr application w configurd th pirnt tcntr with on virtual ntwork for ach virtual machin. hi allowd u to grgat ntwork activity and dirct traffic to pcific rvr and nurd that w prad th load vnly acro all th participating virtual machin. W firt ran om tt to dtrmin how many concurrnt ur a ingl V could handl bfor prformanc bgan to dgrad. Prformanc dgradation wa indicatd by tranaction timing out and incraing rpon tim. W alo monitord X rvr prformanc figur to nur that th ytm did not nd to wap mmory to dik. With narly 2GB of RA availabl and a V that ndd only 1GB w did not xpct to any wapping occurring and thi turnd out to b th ca. 3

4 t Procdur W ud thi figur to plan load on th tt with multipl virtual machin. W cho concurrnt ur rathr than th mor uual tranaction pr cond bcau ach ur followd a t cript that accd variou pag on th wb it. ach imulatd ur would pnd tim on ach pag bfor rtriving th nxt on, and a imulatd ur would abort th tranaction if thy rcivd no rpon within thirty cond. hi activity trd th rvr by tying up rvr mmory for buffr and ion information, which in turn incrad dmand on th hot mmory. For xampl, whn running thr virtual machin with th hot configurd with 2GB of mmory w found that wapping took plac a oon a w activatd th othr two V. Whn w applid th am tt load to all thr ytm th hot mmory bcam ovr committd and HP tranaction bgan to fail. In thory thr V hould hav bn abl to handl thr tim th workload that a ingl V could cop with, but in practic thi provd to b only partly th ca. Although th ytm handld th load impod on it, th rpon tim incrad and om tranaction faild to complt. Avrag rpon tim incrad from 13 m to 198 m. Whn w carrid out th databa tt w ud th am imulatd ur tchniqu a bfor, but in thi ca th wb pag wr ud to triggr variou databa opration uch a rtriving and updating databa row and xcuting complx QL quri. Whn th rqutd opration wa compltd th ur rcivd an acknowldgmnt pag, and th lapd tim indicatd how long th opration had takn. h opration trd mmory in variou way, and th tim takn to complt an opration incrad a th dmand on th mmory incrad. Bcau databa rvr tnd to nd mor mmory w would normally hav rconfigurd th virtual machin with a largr mmory iz for th tt. W rtaind th am 1GB iz that w ud for th wb it tt, partly o that w could hav lik for lik comparion btwn th two t of tt, but mainly bcau w would hav ovr-committd th rvr mmory and caud wapping to occur almot immdiatly. hi would hav ditortd th final rult. inc databa opration ar gnrally mor mmory-intniv than wb it thi impod a furthr train on th ytm and it wa no urpri to find that th ytm could not handl th numbr of imulatd ur that th wb it tt could upport. nc again, w dtrmind th lvl of prformanc that could b obtaind with a ingl V in a imilar way to that ud for th wb rvr tt. X rvr prformanc figur indicat how much tr th ytm mmory i undr. A mall haring valu rlativ to th numbr of V in u indicat a l fficint u of ytm mmory, whil a largr valu indicat a mor fficint u. imilarly, although th balloon figur ought idally to b zro, a rlativly larg figur alo indicat an fficint u of mmory. By contrat a low wap figur i much bttr than a larg on inc it rflct mmory ovr-commitmnt caud by inufficint ytm mmory availabl for th numbr of V in u. hard and balloon mmory i allocatd in ytm mmory whil wappd mmory i tord on dik and thi ha an impact on prformanc inc mmory acc i much fatr than dik. pirnt prformanc figur ar traightforward. Avrag rpon tim indicat th kind of tr th mmory i undr. Long rpon tim indicat that th ytm i buy whil hort rpon tim indicat that th ytm i coping aily. Factor affcting rpon tim includ mmory commitmnt and procor u, although om procor load i du to X uing procor tim to manipulat mmory. Faild tranaction ar an indication of a ytm bcoming ovrloadd, and whil mall prcntag figur ar nothing to b concrnd about, high figur indicat a ytm coming undr tr. A tranaction conit of a rqut and an acknowldgmnt that it ha bn rcivd. If th rqut i for om kind of data uch a a wb pag thn th data will b nt latr a a parat tranaction. ranaction may fail for a numbr of raon but th ffct i th am in that a rqut i nt to a rvr and no acknowldgmnt i rcivd within th xpctd tim limit. 4

5 rvr Rpon im h tt rult how clarly th impact th amount of phyical mmory can hav on th prformanc of Vwar V running wb and backnd databa rvr. With th rvr fittd with 2GB of mmory it wa not poibl to run tt with mor than tn V a w wr unabl to load any mor. Wb rvr rpon to clint rqut acro all combination of V and imultanou ur wr handld in a timly mannr with 32GB of mmory but rpon tim dgradd ignificantly whn accing th backnd databa. Wb rvr Avrag Rpon im i l l i c o n d GB 32GB 1 / / / / / / 55 V / imultanou Ur ataba rvr Avrag Rpon im i l l i c o n d GB 32GB 1 / 4 3 / 12 6 / 24 1 / 4 15 / 6 2 / 8 V / imultanou Ur 5

6 Wb rvr mory Uag For th tt with 2GB, it i clar that Vwar mmory managmnt had it work cut out. A th V/ur load incrad, ballooning bcam mor prvalnt a did pag wapping to dik. hi xplain th rapidly dtriorating prformanc of th wb and databa rvr whn mmory bcam ovrcommittd. uring all tt with 32GB of mmory no ballooning or pag wapping occurrd and th graph how a conitnt ri in hard mmory a th load wa incrad. Wb rvr mory Uag 2GB B y t hard mory (B) Balloon mory (B) wap mory (B) 1 / / / / / / 55 V / imultanou Ur Wb rvr mory Uag 32GB B y t hard mory (B) Balloon mory (B) wap mory (B) 1 / / / / / / 55 V / imultanou Ur 6

7 ataba rvr mory Uag h load placd on th V during th databa tt ar much highr than for th wb rvr tt. ach ur bing imulatd by th tcntr i rquting information from th wb rvr which mut, in turn, rtriv thi from th aociatd QL backnd databa. A can b n in th 2GB tt, mmory haring incra rapidly for th firt two load and thn ballooning and pag wapping to dik cam into play a mmory bcam riouly ovrcommittd. A with th wb rvr tt, 32GB of mmory wa ufficint at th load to allow mmory haring only to b rquird. ataba rvr mory Uag 2GB B y t hard mory (B) Balloon mory (B) wap mory (B) 1 / 4 3 / 12 6 / 24 1 / 4 15 / 6 2 / 8 V / imultanou Ur ataba rvr mory Uag 32GB B y t hard mory (B) Balloon mory (B) wap mory (B) 1 / 4 3 / 12 6 / 24 1 / 4 15 / 6 2 / 8 V / imultanou Ur 7

8 rvr Faild ranaction Wb rvr tt with 2GB of mmory how that faild tranaction incrad dramatically to 65 pr cnt. h rvr with 32GB of mmory only faild to rpond to l than on pr cnt at th hight load. f mor intrt i th failur rat for th databa tt a with only 2GB of mmory thi rachd an unaccptabl 69 pr cnt with only thr V loadd. Howvr, vn with 32GB of mmory th V wr incraingly unabl to rpond to ur rqut for data a th load rampd up. hi how clarly that conidrably mor phyical mmory would b rquird to rvic thi numbr of V and ur. Wb rvr Faild ranaction % GB 32GB 1 / / / / / / 55 V / imultanou Ur ataba rvr Faild ranaction % GB 32GB 1 / 4 3 / 12 6 / 24 1 / 4 15 / 6 2 / 8 V / imultanou Ur 8

9 Concluion h rult from th tt clarly dmontrat th ffct th amount of phyical mmory can hav on th prformanc of V running undr Vwar X rvr. With 2GB of phyical mmory a rqutd by Kington chnology w immdiatly ncountrd problm a th rvr wa unabl to upport mor than tn V. A w importd V w found th ach ubqunt job took longr and th lvnth import job faild to complt. Conquntly, th job rquiring fiftn and twnty V hav bn markd in th rlvant graph a not ttd. Wb rvr rpon tim for 2GB of mmory bcam progrivly wor and for tn V with a load of 27,5 ur it actually rachd a high of narly ightn cond. For th databa rvr tt with 2GB thi wa far wor with tn V and 4 imultanou ur having to put up with avrag rpon tim of narly on minut. From th X rvr mmory uag tatitic w concludd that for 2GB th only configuration th rvr could upport without any mmory managmnt coming into play wa a ingl V. Anything byond thi and ballooning wa initiatd and, wor till, pag wapping to dik. Upgrading th rvr to 32GB of mmory a rqutd by Kington chnology allowd u to run our tt acro th full rang of V. Howvr, although rpon tim for th wb rvr nvr wnt abov two cond, th rult for th databa rvr tt wr unaccptably high. With twnty V and ighty imultanou ur, th tim for th rvr to rpond climbd to an avrag of ovr forty cond. X rvr mmory managmnt had l to do with 32GB in th rvr with it rporting mmory haring a th only activity occurring. For all tt thr wa no vidnc of ballooning or pag wapping bing rquird. Although th 32GB of mmory appard to b handling th tt wll th numbr of faild tranaction a rportd by th tcntr wa quit rvaling. h ar imulatd ur rqut that do not rciv any acknowldgmnt from th rvr thy wr dirctd to. n th wb rvr tt th prcntag of faild tranaction for 32GB nvr ro abov.5 pr cnt. Howvr, it faild to handl th highr dmand of rqut dirctd to th databa rvr with th final tt uing twnty V and ighty imultanou ur rporting an avrag numbr of faild tranaction of 57 pr cnt. any buin ar moving to virtualiation to rduc oprational cot but a failur to configur thir rvr corrctly will hav a ngativ impact on prformanc. Although tting wa limitd to rvr configuration with 2GB and 32GB of mmory, th rult clarly how th ubtantial prformanc improvmnt that can b mad for V running wb rvr and databa rvic by upgrading mmory. Copyright 29 Binary ting Limitd. All right rrvd. o part of thi publication may b rproducd in any form without prmiion in writing from Binary ting Limitd. Binary ting Limitd bliv th information in thi publication to b accurat a of it publication dat and rtain th right to mak chang to thi publication at any tim without notification. Binary ting Limitd mak no warranty with rpct to th information that appar in thi publication, aum no rponibility for any rror that may appar and do not mak a commitmnt to updat th information containd thrin. All tradmark rfrncd in thi publication blong to thir rpctiv compani 9