USF Libraries
USF Digital Collections

Power-performance tradeoff in database systems

MISSING IMAGE

Material Information

Title:
Power-performance tradeoff in database systems
Physical Description:
Book
Language:
English
Creator:
Xu, Zichen
Publisher:
University of South Florida
Place of Publication:
Tampa, Fla
Publication Date:

Subjects

Subjects / Keywords:
Database management system
Power modeling
Power estimation
Energy concern
Query Opotimization
Dissertations, Academic -- Computer Science -- Masters -- USF   ( lcsh )
Genre:
non-fiction   ( marcgt )

Notes

Abstract:
ABSTRACT: With the total energy consumption of computing systems increasing at a steep rate, much attention had been paid to the design of energy-efficient computing systems and applications. So far, database system design has focused on improving the performance of query processing. The objective of this study is to explore the potential of energy conservation in relational database management systems. The hypothesis is: by modifying the query optimizer in a Database management system (DBMS) to take the energy cost of query plans into consideration, we will be able to reduce the energy usage of database servers and control the tradeoffs between energy consumption and system performance. In this thesis, we provide an in-depth anatomy of typical queries in various benchmarks and qualitatively analyze the energy profile of such queries. The results of extensive experiments show that power savings in the range of 11% to 22% can be achieved by equipping the DBMS with a simple query optimizer that selects query plans based on both estimated processing time and energy requirements. We advocate more research efforts be invested into the design and evaluation of power-aware DBMSs in hope to reach higher level of energy efficiency.
Thesis:
Thesis (M.S.C.S.)--University of South Florida, 2009.
Bibliography:
Includes bibliographical references.
System Details:
Mode of access: World Wide Web.
System Details:
System requirements: World Wide Web browser and PDF reader.
Statement of Responsibility:
by Zichen Xu.
General Note:
Title from PDF of title page.
General Note:
Document formatted into pages; contains 53 pages.

Record Information

Source Institution:
University of South Florida Library
Holding Location:
University of South Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
aleph - 002068290
oclc - 606604083
usfldc doi - E14-SFE0003124
usfldc handle - e14.3124
System ID:
SFS0027440:00001


This item is only available as the following downloads:


Full Text
xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam 2200385Ka 4500
controlfield tag 001 002068290
005 20100412141110.0
007 cr mnu|||uuuuu
008 100412s2009 flu s 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0003124
035
(OCoLC)606604083
040
FHM
c FHM
049
FHMM
090
QA76 (Online)
1 100
Xu, Zichen.
0 245
Power-performance tradeoff in database systems
h [electronic resource] /
by Zichen Xu.
260
[Tampa, Fla] :
b University of South Florida,
2009.
500
Title from PDF of title page.
Document formatted into pages; contains 53 pages.
502
Thesis (M.S.C.S.)--University of South Florida, 2009.
504
Includes bibliographical references.
516
Text (Electronic thesis) in PDF format.
3 520
ABSTRACT: With the total energy consumption of computing systems increasing at a steep rate, much attention had been paid to the design of energy-efficient computing systems and applications. So far, database system design has focused on improving the performance of query processing. The objective of this study is to explore the potential of energy conservation in relational database management systems. The hypothesis is: by modifying the query optimizer in a Database management system (DBMS) to take the energy cost of query plans into consideration, we will be able to reduce the energy usage of database servers and control the tradeoffs between energy consumption and system performance. In this thesis, we provide an in-depth anatomy of typical queries in various benchmarks and qualitatively analyze the energy profile of such queries. The results of extensive experiments show that power savings in the range of 11% to 22% can be achieved by equipping the DBMS with a simple query optimizer that selects query plans based on both estimated processing time and energy requirements. We advocate more research efforts be invested into the design and evaluation of power-aware DBMSs in hope to reach higher level of energy efficiency.
538
Mode of access: World Wide Web.
System requirements: World Wide Web browser and PDF reader.
590
Advisor: Yicheng Tu, Ph.D.
653
Database management system
Power modeling
Power estimation
Energy concern
Query Opotimization
690
Dissertations, Academic
z USF
x Computer Science
Masters.
773
t USF Electronic Theses and Dissertations.
4 856
u http://digital.lib.usf.edu/?e14.3124



PAGE 4

LISTOFFIGURESiii ABSTRACTiv CHAPTER1INTRODUCTION1 CHAPTER2BACKGROUND5 2.1DatabaseConceptandQueryOptimization5 2.2BenchmarkingEnergySavings9 2.3RelatedWorkinArchitectureLevel12 2.4RelatedWorkinSystemLevel14 2.5RelatedWorkinSoftwareLevel15 CHAPTER3HYPOTHESISANDMOTIVATINGEXAMPLES18 3.1RationaleBehindtheHypothesis19 3.2AMotivatingExample20 CHAPTER4METHODOLOGY24 4.1PowerModel25 4.2PlanEvaluationModel27 4.3PowerModelCalibration28 4.4Workload29 4.5ExperimentalDesign30 4.6TestbedDescription31 CHAPTER5EXPERIMENTALRESULTS32 5.1MainResults(Task1)32 5.2QueryCharacterization(Task2)37 5.3MultipleWorkloads(Task3)39 5.4Discussions39 CHAPTER6CONCLUSIONSANDFUTUREWORK45 REFERENCES46i

PAGE 5

Table3.2PowerconsumptionofvariousSeagateMomentusharddisks.19 Table3.3Powercostsofbasicdatabaseoperationsintheexperimentalsystem.23 Table4.1Hardwarepowerspecicationsintheexperimentaldatabaseserver.26 Table4.2Powercostfunctionsforaccessingsinglerelationsandjoinoperations.27 Table4.3Powercostsofthreesingle-tablescanoperationsinmodelvericationex-periments.29 Table5.1Performanceandpower/energyconsumptionoftheexperimentsusingTPC-HworkloadandTPC-Cworkload.36ii

PAGE 6

Figure3.2TheestimatedprocessingtimeandpowercostofplansevaluatedbythemodiedPostgreSQLqueryoptimizerforthequeryshowningure3.1.22 Figure3.3Apartialplantreeofthequeryingure3.2generatedbythemodiedqueryoptimizer.23 Figure4.1Organizationoftheexperimentplatform.25 Figure5.1PowerconsumptionoftheTPC-Hworkloadsunderthreedierentdatabasesizes.33 Figure5.2PowerconsumptionoftheTPC-Cworkloadsunderthreedierentdatabasesizes.34 Figure5.3Powerandprocessingtimeofthechosenplansforthe19queriesintheTPC-Hbenchmarkasestimatedbythequeryoptimizer.41 Figure5.4Power,processingtime,andenergyconsumptionofthe19queriesintheTPC-Hbenchmarkasmeasuredbypowermeter.42 Figure5.5Relativequantityofpower.43 Figure5.6Absolutequantityofpower.44iii

PAGE 8

Whiletheabovecostsarecalculateddirectlyfromenergyconsumption,power(i.e.,energyconsumptionperunittime)savingsareofmorepracticalimportancethanenergysavingsinsystemdesign.Powercappinghasbecomeaseriouschallengefordatacentersinrecentyears.Controllingpowerconsumptionisanessentialwaytoavoidsystemfailurescausedbypowercapacityoverloadoroverheatingduetoincreasinghighserverdensity(e.g.,bladeservers)[27,55].Modernhigh-densityserversfaceanincreasingprobabilityofthermalfailuresduetotheircontinuouslydecreasingsizeandincreasingdemandforcomputationalcapabilities.Forexample,recentstudiesshowthat50%ofallelectronicsfailuresarerelatedtooverheating[58].Inaddition,anapproximately15Cincreaseintemperaturecoulddoublethefailurerateofadiskdrive[4].Therefore,itisimportanttoreducedatabasepowerconsumptionsothatthetotalpowerconsumptionofanentiresystemcanbekeptbelowagivenpowerbudget.Furthermore,1

PAGE 9

Duetotheabovefacts,powerconservationincomputingsystemshasattractedmuchat-tentionfromgovernmentagencies,industrialsectors,andresearchcommunities.Standardsandbenchmarksarequicklyadaptingtotheconsiderationofpowerconsumption[46,1,50].Intheresearchpart,earlyinnovationsonpower-awarecomputerservershaveconcentratedonhardwareandsystemsoftwaresuchascompilersandoperatingsystems[35,59,41,20,10].Thosebuildafoundationforacomputingenvironmentwhere:1)hardwarecanoperateonvariouspower-savingmodesinwhichdierentpower/performancetradeoscanbeachieved;2)systemsoftwarecancontroltheoperatingmodesofhardwareandrescheduleresourcerequestsfromapplicationstosaveenergy.Onecommonthemeofsystem-levelresearchonpower-awarecomputingisthattheytreatthehigh-levelapplicationsaspassiveresourceconsumers(ab-stractedasworkload)thatarenotawareofthelow-leveleortsofpowerreduction.Inthepastfewyears,theresearchcommunityhasshiftedmuchinteresttopower-awareapplications[9,57,13],whichprovidesignicantsynergisticvaluestocurrentresearchonthehardwareandsystemlevels.Ourresearchfallsintothiscategory. Inthisthesis,westudypowerconsumptionpatternsandidentifypower-savingopportunitiesindatabasemanagementsystems(DBMS).Ourultimategoalistobuildpower-awareDBMSsthatcanachievesignicantcostsavingswhilemaintainingreasonableperformanceinqueryprocessing.Suchvisionismotivatedbythefollowingobservations. First,powerreductioninDBMSsisofhigheconomicalsignicance:DBMSisanimportanttypeofsoftwareinthethree-tiercomputingarchitectureadoptedbymostoftoday'sbusinesscomputingenvironments.Inatypicaldatacenter,amajorityofthecomputingresourcesarededicatedtodatabaseservers,makingDBMSthelargestconsumersofpowerinallthesoftwareapplicationsdeployed.Itisgenerallytruethattheprocessingcapacityoftheback-enddatabaseserversdeterminesthatofthefront-endwebandapplicationservers.In[48],aback-endtofront-endpowerconsumptionratioof11:9:9isreported.Mostdatabaseserversareconguredwithcapacitysucienttohandlepeakload.Thisimpliesopportunitiesforpowersavings.2

PAGE 10

Themainobjectiveofthisstudyistovalidateourvisiononpower-awareDBMSs(PDBMS)andprovidesolutionstosomeofthemainproblemsrelatedtobuildingsuchsystems.Speci-cally,wewanttoanswerthefollowingquestions:

PAGE 11

Thisthesisstatesthatitispossibletosavepower,orevenenergyinmanipulatingoptimizerindatabasesystemwithfullsupportofexperimentalresult.Inanotherword,thereispotentialthatmanipulatingsoftwareitselftoreachapromisingsavinginpower.Also,itprovidesafreescratchskeletonoftestingaP-DBMSwithvericationofitspower-estimationdatamodel.Atlast,thethesispointsoutseveralpossibledirectionsofpowersavinginDBMSinfurtherresearch. Thethesisisorganizedasfollows:relatedworksummarizedinchapter2;ourhypothesisisstatedandsoaremotivatingexamplesinchapter3;chapter4describesthemethodologyandenvironmentforourexperiments;wepresentandinterprettheexperimentalresultsinchapter5andconcludethisthesiswithdiscussionsonfutureresearchplansinchapter6.4

PAGE 12

Dependingontheintendedpurpose,thereareanumberofdatabasearchitecturesinuse.Manydatabasesuseacombinationofstrategies.TherearetwomainbasisDBMSarchitec-tures.Oneison-lineTransactionProcessingsystems(OLTP)oftenusearow-orienteddatastorearchitecture,whiledata-warehouseandotherretrieval-focusedapplicationslikeGoogle'sBigTable[19],orbibliographicdatabase(librarycatalogue)systemsmayuseaColumn-oriented[53]DBMSarchitecture. Document-Oriented,knowledgebases,XML,aswellasframedatabasesandRDF-stores(asknownastriple-stores),alsouseacombinationofthesearchitecturesintheirimplementation.Itshouldbenotedthatnotalldatabaseshaveorneedadatabase'schema'(socalledschema-lessdatabases).OvermanyyearsthedatabaseindustryhasbeendominatedbyGeneralPurposedatabasesystems,whichoerawiderangeoffunctionsthatareapplicabletomany,ifnotmostcircumstancesinmoderndataprocessing.Thesehavebeenenhancedwithextensibledatatypes,pioneeredinthePostgreSQLproject1,toallowaverywiderangeofapplicationstobedeveloped.Inourexperiments,PostgreSQLisusedasourtestbenchtocertifyourhypothesis,verifyourpower/energymodelandimplementit.

PAGE 13

MostrelationalDBMS'sandsomeobjectDBMSshavetheadvantagethatindexescanbecreatedordroppedwithoutchangingexistingapplicationsmakinguseofit.Thedatabasechoosesbetweenmanydierentstrategiesbasedonwhichoneitestimateswillrunthefastest.Inotherwords,indexesaretransparenttotheapplicationorend-userqueryingthedatabase;whiletheyaectperformance,anySQLcommandwillrunwithorwithoutindextocomputetheresultofanSQLstatement.TherationalDBMS(RDBMS)willproduceaplanofhowtoexecutethequery,whichisgeneratedbyanalyzingtheruntimesofthedierentalgorithmsandselectingthequickest.Someofthekeyalgorithmsthatdealwithjoinsarenestedloopjoin,sort-mergejoinandhashjoin.Whichoftheseischosendependsonwhetheranindexexists,whattypeitis,anditscardinality. Queryoptimization2isafunctionofmanyRDBMSsinwhichmultiplequeryplansforsatisfyingaqueryareexaminedandagoodqueryplanisidentiedasdiscussedintheindexdescriptionabove.Thereisnoonehundredpercentguaranteeforaabsolutebeststrategybecausetherearemanywaysofcreatingplans.Thersttradeoisbetweentheamountoftimespentguringoutthebestplanandtheamountrunningtheplan.Dierentqualitiesofdatabasemanagementsystemshavedierentwaystobalancethesetwo.InPostgreSQL,costbasedqueryoptimizersevaluatetheresourcefootprintofvariousqueryplansandusethisasthebasisforplanselection. TypicallytheresourceswhicharecostedareCPUpathlength(CPUtuples),amountofdiskbuerspace(usedpages)andinterconnectusagebetweenunitsofparallelism(additionalcost). optimizer6

PAGE 14

Thegoalofoptimizationistoeliminateasmanyunneededtuples,orrowsaspossible.Thefollowingisalookatrelationalalgebraasiteliminatesunneededtuples.Theprojectoperatorisstraightforwardtoimplementifinchapter3containsakeytorelationR.IfitdoesnotincludeakeyofR,itmustbeeliminated.Thismustbedonebysorting(seesortmethodsbelow)andeliminatingduplicates.ThismethodcanalsobedonebyhashingtoeliminateduplicatesHashtable.ThenSQLcommanddistinctisconsidered;thisdoesnotchangetheactualdata.Thisjusteliminatestheduplicatesfromtheresults.Moreover,operationssetsarecollected.Databasemanagementheavilyreliesonthemathematicalprinciplesofsettheorywhichiskeyincomprehendingtheseoperations. Union,intersectionandsetdierencedisplayallthatappearinbothsets,eachlistedonce.Forintersection,itismorerestrictedthatitonlylistsitemswhosekeysappearinbothlists.Nevertheless,setdierencelistsallitemswhosekeysappearintherstlistbutnotthesecondone.Thesethreemustbeunioncompatiblewhichmeansallsequencesofselectedcolumnsmustdesignatethesamenumberofcolumns.Thedatatypesofthecorrespondingcolumnsmusttherebycomplywiththeconditionsvalidforcomparability.Eachdatatypecanbecomparedtoitself.Forexamples,columnsofdatatypeCHARwiththedierentASCIIandEBCDICcodeattributescanbecomparedtoeachother,wherebytheyareimplicitlyadapted;columnswiththeASCIIcodeattributecanbecomparedtodate,time,andtimestampspecications;allnumberscanbecomparedtoeachother.ThelastisCartesianProduct,suchas(R*S),whichtakesalotofmemorybecauseitsresultcontainsarecordforeachcombinationofrecordsfromRandS.7

PAGE 15

InPostgreSQL,itfollowsthesamemechanism.Foreachplanningproblem,therefore,therewillbealistofrelationsthatareeitherbaserelationsorjoinrelationsconstructedpersub-join-lists.Theserelationsjoinedtogetherinanyordertheplannerseest.Thestandardplannerdoesthisasfollows: R/8

PAGE 17

TheTransactionProcessingPerformanceCouncil(TPC)5isanon-protcorporationfoundedtodenetransactionprocessinganddatabasebenchmarksandtodisseminateobjective,ver-iableTPCperformancedatatotheindustry.Theterm\transaction"isoftenappliedtoawidevarietyofbusinessandcomputerfunctions.WhilstTPCbenchmarkscertainlyinvolvethemeasurementandevaluationofcomputerfunctionsandoperationsasSPEC,anditinvolvesdatabasewhichismytarget.Thus,thisbenchmarkisusedinmypreviousworkregardsasbenchmark.TheTPCregardsatransactionasitiscommonlyunderstoodinthebusinessworld:acommercialexchangeofgoods,services,ormoney.Atypicaltransaction,asdenedbytheTPC,wouldincludetheupdatingtoadatabasesystemforsuchthingsasinventorycontrol(goods),airlinereservations(services),orbanking(money).Intheseenvironments,anumberofvirtual\customers"or\service"representativesinputandmanagetheirtransactionsviaaterminalordesktopcomputerconnectedtoadatabase.Typically,theTPCproducesbench-marksthatmeasuretransactionprocessing(TP)anddatabase(DB)performanceintermsofhowmanytransactionsagivensystemanddatabasecanperformperunitoftime,e.g.,trans-actionspersecondortransactionsperminute.Inthiskindofsimulation,thereareperfectlydierentkindsofqueriesandrequestwhichwillconsumedinlargedatabasesystem.Withthoseinformation,thehypothesisinchapter3canbeveried. Morespecically,TPCisabouttoapprovethethirdinitsseriesofbenchmarkswhichmeasuretheperformanceandprice/performanceoftransactionprocessingsystems.Forex-ample,thenewTPCBenchmarkC,orTPC-C,isanon-linetransactionprocessing(OLTP)benchmark.TPC-Ccontainsmultipletransactiontypes,morecomplexdatabase,andoverallexecutionstructure.TPC-CisbasedonaworkloadpresentedtotheTPCtwoyearsagoandrenedrepresentingacrosssectionoftheindustry. ThegoalofTPC-Cbenchmarksistodeneasetoffunctionalrequirementsthatcanberunonanytransactionprocessingsystem,regardlessofhardwareoroperatingsystem.Itisthenuptothetestsponsortosubmitproofthattheyhavemetalltherequirements.This

PAGE 18

TPCbenchmarksalsodierfromotherbenchmarksinthatTPCbenchmarksaremodeledafteractualproductionapplicationsandenvironmentsratherthanstand-alonecomputertestswhichmaynotevaluatekeyperformancefactorslikeuserinterface,communications,diskI/Os,datastorage,andbackupandrecovery.AsanOLTPsystembenchmark,TPC-Csimulatesacompleteenvironmentwhereapopulationofterminaloperatorsexecutestransactionsagainstadatabase.Thebenchmarkiscenteredaroundtheprincipalactivities(transactions)ofanorder-entryenvironment.Thesetransactionsincludeenteringanddeliveringorders,recordingpayments,checkingthestatusoforders,andmonitoringthelevelofstockatthewarehouses.However,itshouldbestressedthatitisnottheintentofTPC-CtospecifyhowtobestimplementanOrder-Entrysystem.Whilethebenchmarkportraystheactivityofawholesalesupplier,TPC-Cisnotlimitedtotheactivityofanyparticularbusinesssegment,but,rather,representsanyindustrythatmustmanage,sell,ordistributeaproductorservice. IntheTPC-Cbusinessmodel,awholesalepartssupplier(calledtheCompanybelow)oper-atesoutofanumberofwarehousesandtheirassociatedsalesdistricts.TheTPCbenchmarkisdesignedtoscalejustastheCompanyexpandsandnewwarehousesarecreated.However,cer-tainconsistentrequirementsmustbemaintainedasthebenchmarkisscaled.EachwarehouseintheTPC-Cmodelmustsupplytensalesdistricts,andeachdistrictservesthreethousandcustomers.Anoperatorfromasalesdistrictcanselect,atanytime,oneoftheveoperationsortransactionsoeredbytheCompany'sorder-entrysystem.Likethetransactionsthemselves,thefrequencyoftheindividualtransactionsaremodeledafterrealisticscenarios. Themostfrequenttransactionconsistsofenteringaneworderwhich,onaverage,iscom-prisedoftendierentitems.Eachwarehousetriestomaintainstockforthe100,000itemsintheCompany'scatalogandllordersfromthatstock.However,inreality,onewarehousewillprobablynothaveallthepartsrequiredtolleveryorder.Therefore,TPC-Crequires11

PAGE 19

TheTPCBenchmarkH(TPC-H)isadecisionsupportbenchmark.Itconsistsofasuiteofbusinessorientedad-hocqueriesandconcurrentdatamodications.Thequeriesandthedatapopulatingthedatabasehavebeenchosentohavebroadindustry-widerelevance.Thisbenchmarkillustratesdecisionsupportsystemsthatexaminelargevolumesofdata,executequerieswithahighdegreeofcomplexity,andgiveanswerstocriticalbusinessquestions.TheperformancemetricreportedbyTPC-HiscalledtheTPC-HCompositeQuery-per-HourPer-formanceMetric,andreectsmultipleaspectsofthecapabilityofthesystemtoprocessqueries.Theseaspectsincludetheselecteddatabasesizeagainstwhichthequeriesareexecuted,thequeryprocessingpowerwhenqueriesaresubmittedbyasinglestream,andthequerythrough-putwhenqueriesaresubmittedbymultipleconcurrentusers.MoredetailsabouthowTPCbenchmarksinvolveinthetestareinchapter4.2.3RelatedWorkinArchitectureLevel

PAGE 20

Diggingdeeperinmicroarchitectureframework,[11]certifytheabilitytoestimatepowerconsumptioninanearly-stagedenition.Thereexiststrade-ostudieswhichisakeynewmethodologyenhancement.Asitisstated,\savingpowercanbeexposedviamicroarchitecture-levelmodeling,particularlythroughclock-gatinganddynamicadaptation."Intheworkof[30],itnotonlygivesascalingpower-awaremicroarchitectureandpresentsastrategyforrun-timeprolingtooptimizethecongurationofamicroprocessordynamically.Inthewayasdescribe,savingpowerwithminimumperformancepenaltyisachieved.Afterchangesinthecongura-tionoftheprocessoraccordingtotheparallelismintherunningprogram,theworkobservedadecreaseofupto23%inenergy/cycleandupto8%inenergyperinstruction.Ratherthanthestrategyandframework,theydonotfocusthermalissue,anotherwork[52]comesout.Ob-viously,designingcoolingsolutionsforworst-casepowerdissipationisprohibitivelyexpensive.Chipsthatcanautonomouslymodifytheirexecutionandpower-dissipationcharacteristicsper-mittheuseoflower-costcoolingsolutionswhilestillguaranteeingsafetemperatureregulation.Itintroducesseveraleectivewaysfordynamicthermalmanagement(DTM):\temperature-13

PAGE 21

Toimplementsuchtechnique,schedulingworkcomesastherstworkeldthatresearchersfocuseson.Especiallyintherealtimeembeddedsystemwhoconcernsperformanceandpowerasbothcriticalmission.In[39],anewschedulingtechniqueforsupportingthedesignandevaluationofaclassofpower-awaresystemsinmission-criticalapplicationsisestablished.Itsatisesconstraintsofstringentmin/maxtimingandmaxpowerandalsomakesthebesteorttosatisfytheminpowerconstraintinanattempttofullyutilizefreepowerortocontrolpowerjitter.Anotherworkonthesimilartopicisin[38],itstatesthat\power-awaresystemsarenotjustlow-powerbutmusttracktheirpowersourcesandthechangingpowerandperformance14

PAGE 22

Someanalysisworkonthepowerconsumptionofreal-timeoperatingsystems(RTOSs)isdonein[16],whichisanimportantcomponentofthesystemsoftwarelayer.Thisworkisalsoabasisofmanyotherpaperscitedinsection2.5.ThisworkpresentsthepowerprolesforacommercialRTOSrunningseveralapplicationsonanembeddedsystembasedontheFujitsuSPARCliteprocessor.DemonstrationabouttheRTOScanconsumeasignicantfractionofthesystempowerand,inaddition,impactthepowerconsumedbyothersoftwarecomponents. Asmentionedearlier,muchofthepriorworkhaseitherattemptedtoreducepowercon-sumptionbyimprovingthepower-eciencyofindividualhardwarecomponents[35],orfocusedonsystem-levelpowerandthermalmanagement[10,20,24,41,59,49,56].Thecommonphi-losophybehindtheseis:powerconsumptionshouldbeproportionaltotheloadputonthesystem.Inordertosaveenergy,weneedtoturnthehardwaretopower-savingmodeswhentheworkloadonthesystemislight.Amongthesework,Zengetal.[59]managespowerinoperatingsystemwhichprovidesuswithaninitialpowermodelforpowercostestimationofqueryplans.Inthisthesis,wewillshowthat,giventhesameloadofuserqueries,theDBMScanchoosetohandletheloadinapower-savingmanner.2.5RelatedWorkinSoftwareLevel

PAGE 23

Thepowerconsumptionindatabaseshasjuststarteddrawingattentionfromtheresearchcommunity.TheClaremontreportondatabaseresearch[2]explicitlystatestheimportanceof\designingpower-awareDBMSsthatlimitenergycostswithoutsacricingscalability...".Tworecentprojectsaddresspowerconsumptionissuesindataprocessingapplicationsanddatacenters:[50]presentsabenchmarkforevaluatingtheenergyeciencyofimplementationsofvarioussortingalgorithms.Duetothelargevolumeofdata,suchalgorithmsareI/OintensiveandthereforetheI/Oaccesspatternshaveprofoundeectsontheenergyeciencyofthealgorithms.[48]reportsextensiveexperimentalresultsonthepowerconsumptionpatternsintypicalcommercialdatabaseservers.Basedontheseresults,itprovidessuggestionsonhowtomakethesystemmorepowerecient.However,thesesuggestionsfocusonutilizingnewhardwareratherthanmodifyingtheDBMSsoftware.Tothebestofourknowledge,ourprojectistherstonethattakesthepathofredesigningtheDBMSkernelforpower-savingpurposes. Averysimilarworkasourcurrentdatabaseenergysavingjobis[42].Itfocusesonsoftwarelevelapplicationtocontrolpowermanagement.Twoalgorithmsofdynamicpowermanagementimplementforcontrollingthepowerstatesofaharddisk.Itcreatesatemplateforsoftware-controlledpowermanagementandexperimentalcomparisonsofmanagementalgorithmsforaharddisk. Whilelimitingpowerconsumptionisobviouslynecessaryinmobilecomputingsystems(e.g.,PDAs,laptops)inwhichbatterypoweristhemoststringentresource.Thesoftwarecontrolissuesasfollows:transition,load-change,andadaptationproducesbothsoftwareandhardwareproblems.In[40],itimplementsandproposessolutionstosoftwareenergymanagementissuescreatedbyexistingandsuggestedhardwareinnovations. Inall,asstatedin[33],\energy-proportionaldesignswouldenablelargeenergysavingsinservers,potentiallydoublingtheireciencyinreal-lifeuse."Usingenergyusageproleofevery16

PAGE 25

Theabovehypothesisisintuitiveduetothewell-knownfactthatthecurrentqueryop-timizationmechanismsconsiderperformanceasthesolegoal.Therefore,thelackofpowerconsiderationsmakesthemunsuitableforsavingpower.Ontheotherhand,itcouldalsobecounterintuitivetomanysinceitispossiblethatqueryoptimizationtowardsshortestprocess-ingtimecoincideswiththattowardslowestpowercost,andthereforenopowersavingscanbeachievedbymodifyingthebehaviorofthedatabaseengine.Infact,thishasbeenafrequentargumentweencounterinourcommunicationswithfellowresearchers.Theargumentisbasedontheunderstandingofqueryprocessingtimebeingameasureofsome\load"tobenishedbytheDBMS,andmore\load"thesystemfaces,moreenergywillbeconsumed.Mostofthepower-awarecomputingresearchfollowthislineofreasoning.However,morecarefulanalysisofthedierenttypesofresourcesassociatedwithsuch\load"tobehandledintheDBMSwillleadtoadierentconclusion.Table3.1.PeakpowerconsumptionofvariousAMDCPUs. Processor TDP(W) Athlon643500+ 67 Athlon643700+ 89 Athlon64FX-55/57 104 Opteron2.8GHz 93 OpteronDualCore2.2GHzS 93 OpteronDualCore1MBCache642.4GHz 95 OpteronDualCore1MBCache643GHz 98 18

PAGE 26

Disk PeakPower(W) IdlePower(W) 5400.5,320GB,SATA 2:45 0:92 7200.5,320GB,SATA 3:03 0:95 7200.5,100GB,ATA 3:03 0:95 7200.5,160GB,SATA 3:83 1:10 7200.5,100GB,SATA 4:03 1:74

PAGE 27

ToexplorethesearchplansgeneratedbyPostgreSQL,wecarefullymodifyitsplangener-ationprocesssuchthatwecanseeallpossibleplansofaquery.IntheregularPostgreSQLsystem,aheuristicsearchalgorithmthatdropsintermediateplansgreedilyisused,inordertoachievepolynomialtimetraverseofthespacethatgrowsexponentiallywiththenumberoftablesinvolved.However,italsokeepsatriggertoallowexhaustivesearch,andwepullthetriggertocollectdataforallthepossibleplans.TensofthousandsofplanscouldbegeneratedinthebackendPostgreSQLserverwhenaqueryisinpreprocessingstage,alongwiththeirestimatedtimecostandpowercost.ThelatterisrenderedbyourpowermodelthatwillbediscussedinSection4.1.3.2AMotivatingExample

PAGE 28

FromFig.3.2,wecanseethereareatleasttwoplannodesonthenon-dominantfrontier,whichareshownascircleddotsinthemagniedportionofthegraph.Theright-circlednode(ontheright)hasapowercostof5.12andprocessingtimeof5.15whiletheleftnodetoitslefthaspowerandtimecostsof4.16and5.62,respectively.Ifweusetheoriginalqueryoptimizer,therightonewouldhavebeenpickedbecauseofitssmallestimatedtimecost.However,ifwechoosetheleftnode,althoughthetimemayincreaseto5.62,thepowercostdecreasesto4.16{thatistotradea5%performancedegradationfora18:75%powersaving.Thissavingissubstantialiftheestimationcomeswithreasonableprecision.Apparently,suchtradeoscanbeutilizedtoachievepowerreduction.Noteherethatthelargersearchspaceforonequeryis,themorelikelysuchtradeoscanbefound.Ontheotherhand,singletablequeryschedulinghaslimitedpower-savingpotentialinourcurrentmethod. Insummary,webelievethedescriptionsinSection3.1andtheabovemotivatingexampleprovidesananswertoquestion2listedinSection1.Wewillalsoverifyouranswerexperimen-tallyinSection5.2. Inadditiontothecostsoftheaboveinterestingplans,wealsoexploredtheircomputationaldetailsforalow-levelexplanationofthecostdierencesbetweenthem.Themaindierencebetweenthejoinpathsofthesetwoplansis:thelastjoininthegreennodeisahashjoinwhiletheonefortherednodeisamergejoin.InFig.3.3,wepresentsuchdetailsintheformoftheplantreesgeneratedbythePostgreSQLqueryoptimizer.Eachsquareinthegraphrepresentsapartialplan:asingle-tablescan(ontheleaflevel)orjoinofmultipletables(onhigherlevels).Eachsquarecontainstheaccessmethod(algorithm)namefollowedbytheIDsoftablestobejoined,andtheestimatedenergyandtimecostsoftheplanin(x,y)21

PAGE 30

FunctionParameterSymbolDefaultValueMeaning DEFAULT CPU TUPLE0.4EstimatedCPUpowercost POWER COSTpertuple DEFAULT CPU INDEX TUPLE0.05EstimatedCPUpowercost POWER COSTperindexedtuple DEFAULT PAGE POWER4.7EstimatedpowercostforW/R COSTonepagewithoutbuering 23

PAGE 31

ThearchitectureofourexperimentalplatformisshowninFig.4.1.Fortheeaseofrepro-ducingourresults,weprovidemoredetailsofthecomponentsinthisplatformasfollows.

PAGE 33

CPUMemoryDiskTotal Power9894=362:63136.63 Inthisstudy,wealsofollowtheabovestrategytoestimatethepowercostofplans.WerstdeneastaticpowerproleforeachbasicoperationinqueryprocessingandmaintainsuchpowercostsassystemparametersoftheDBMS.Bythis,wecancalculatethepowercostofanentireplangivenitsoperationvector.ThepowerproleofbasicoperationsforourexperimentalsystemisshowninTable3.3.Theinitialvaluesoftheseparameters(datanotshown)wereobtainedfromthespecicationsofthehardwarecomponentswendinourserverasin3.1and3.2anddividedbyrelatedestimatedtimefromPostgreSQL.Forexample,inaCPUofactivepower(i.e.,peakpowerminustheidlepower)20watts,PostgreSQLestimatesthetimetoprocessonenon-indexedtupletobe0.04ms,theestimatedCPUpowerpertuplewouldbe20Wtimes0.04msanddividedbyonesecond{theresultis0.75mW.Thepowerestimationfordiskandmemoryfollowsthesameway. Moreover,incalculatingthepowercosts,weassume(asdidin[48])\thepeakpowercon-sumptionofanentiresystemduringthemeasurementintervalisidenticaltotheaggregateoftheindividualnameplatepowerconsumption".Asformainmemory,theapproximatepowercon-sumptionissetto9wattsperDIMMmodule,assuggestedby[18].ThepowerconsumptionofdiskdrivesandCPUareobtainedfromthemanufacturers'websites.ThepowerspecicationsofthemainhardwarecomponentsinourexperimentalsystemarelistedinTable4.1. Notethatthehardwarespecicationsareprovidedbythehardwarevendorsandarebest-eortestimationsoftherealparameters.Inordertogetmorerealisticestimations,weruncalibrationteststogetthenalper-operationpowercostslistedinTable3.3.Moredetailsaboutmodelcalibrationwillbepresentedin4.3. Thepowercostofaplaniscalculatedfromthoseofthehigher-leveloperations,whichconsistofbasicoperationsshowninTable3.3.Table4.2liststheformulaeforthecomputationofthepowercostofsingle-relationscansviadierentaccessmethodsandtwo-tablejoins.Clearly,theseformulaeusethevaluesofbasicoperationsasbuildingblocks.Again,wefollowtheexactsamemechanismforcalculatingtimecostsinPostgreSQLtogeneratetheseformulae.26

PAGE 34

Methods Costfunction SeqScan +nP(innerpaths) Merge +P(innerpaths)) Hash Specically,ametricmodelofthefollowingformatisadopted:C=PTn=ETn1(4.1) whereCistheaggregatedcostandnisacoecientthatreectstherelativeimportanceofPandT.Intuitively,itmeanswearewillingtosacriceadn-timedegradationinperformancetoachievead-timepowerreduction.Themodelisgeneralinthatwecanbeusedfordierentoptimizationgoalswiththechoiceofn.Whenn=1,weonlyconsiderthetimecost(i.e.,asintheoriginalPostgreSQLdoes.Inpractice,wesimplydisregardPbysettingC=T.);forn=0,weoptimizetowardslowestpowerconsumption;andforn=1,powerandtimeperformancearebothtakenintoconsideration{thecostofaplanisbasicallyitsenergyconsumption.27

PAGE 35

Weuseworkloadsconsistingofsimplequerieswhoseexecutionplansuseonlyoneofthefollowingtablescanmethods:sequentialscan,indexscanandbitmapscan.Thecalibrationprocedureisasfollows:

PAGE 36

EstimatedMeasureDierence Power(w)Power(w) SeqScan24.120.27.2% IndexScan35.230.814.5% BitmapScan30.528.18.5% comparedwiththosemeasuredbypowermeterinascan-onlyqueryprocessingexperiment.Allthequeriesintheseexperimentsareexecutedbyasingletypeofbasictablescanmethods,withoutanyotherscanmethodsorjoinoperationsbecauseitisdiculttotellthepowerconsumptionofdierentoperationsinacomplexqueryplan.Foreachexperimentinvolvingonebasicscanmethod,wefeedthePostgreSQLwith100suchconcurrentqueriesandrecordthepowerconsumptionofthewholeserver.6TheestimatedvaluesarecalculatedfromformulaelistedinTable4.2withtheTandNvaluedobtainedfromthePostgreSQLoptimizer.Table4.3showsthedierencebetweenpowermodelestimationusingthecalibratedparametersandactualpeakpowerwhenexecutingthesesimplequeries.Itisclearthatthepowermodeloverestimatesthetotalpowerconsumptionintheserverforsingleplanprocessing.However,theseestimationsareconsideredcloseenoughtotherealvalues.Thedierencevariesfrom7.2%to14.5%{theseareverypromisingresultsascomparedto18%reportedin[48].4.4Workload Intheexperiment,wechoseTPC-Hasourmaintestingbenchmark.TheTPC-Hbenchmarkillustratesdecisionsupportsystemsthatexaminelargevolumesofdataandexecutes22dierenttypesofquerieswithahighdegreeofcomplexity.TwosalientfeaturesoftheTPC-Hbenchmark

PAGE 37

Thedetailedprotocolsfortheabovetaskswillbedescribedwhenwepresenttheexperi-mentalresultsinSection5. WerunallourexperimentsinasingledatabaseserverequippedwithoneDual-coreAMDOpteronProcessor(type2222SE),4GBofmainmemory,andasingle135GBharddisk.The

PAGE 38

Thekernelofoperatingsystemismodiedtoeliminatetheeectofbuerreadandbuerwriteinpageswitchingandlesystembackup.Inthisway,itwillreducetheeectofpowerconsumptionofthosekindsofoperations.Also,thebackendofPostgreSQLiscarefullyrepro-grammed,recompiledandrebuilttofulllthreepurposes:1.implementingallthedatamodelsdescribedinsection4.2and4.1;2.increasingthecapacityoftheserverendofPostgreSQLsothatitcanprocessmoreclientsrequestorhaltthemsimultaneously;3.recongureitselftotthemodiedoperatingsystemplatform.31

PAGE 39

InFig.5.1and5.2,weplottheinstantaneouspowerconsumptionoftheworkloadintherst800secondsofeachexperiment.1Wecanclearlyseethat,systemrunsonsignicantlylowerpowerwhenwechooseaqueryevaluationmodelthatfavorslow-powerplans.Whenwecomparetheenergy-only(n=0)withtheperformance-only(n=1)results,weobservea

PAGE 40

ThepowervaluesreportedinTable5.1aretheaveragepowerovertheentireworkloadprocessingperiod.Onemightarguethatthisisnotaccuratebecausewemayhavemissedpowerspikesundertheone-secondsamplingperiod.However,byreadingthedata,webelievethisshouldnotbeaconcern.Notethat,inallcases,themeasuredpowerlevelisverystableovertimewithoutanysignicantuctuations.Wecouldhavemissedsomepowerspikes,butitisveryunlikelytomissallofthem.Infact,thisisalsotheresultofourexperimentaldesign:the33

PAGE 41

Anothersideofthestoryistheperformancedegradationofpower-awareDBMSs.InTable5.1,weshowthetotalqueryprocessingtimeoftheTPC-Hexperiments.Itisobviousthatittookmoretimeforthepower-awaresystemstonishprocessingallthequeries,butthedierencesaresmallascomparedtothedierencesofpower.Asaresult,westillyieldtotalenergysavingsof3to7%. Figure5.2.PowerconsumptionoftheTPC-Cworkloadsunderthreedierentdatabasesizes.34

PAGE 42

AninterestingresultcanbeseenfromtheTPC-Cexperimentwiththe5GBdatabase:theactualqueryprocessingtimeunderthepower-awarequeryoptimizer(120mins)isevenlowerthanthatundertheperformance-drivenqueryoptimizer(123mins).Thisisduetothewell-knownfactthatPostgreSQLdoesnotguaranteetheselectionoftheoptimalplan,asaresultoftheestimationerrorsimposedbyitstimecostmodel.Actually,wealsoobservedthisphenomenoninanumberofindividualqueriesintheTPC-Hexperiments:thepower-awareoptimizerfoundplansthataresuperiorinbothpowerconsumptionandactualprocessingtime.ItisonlyintheTPC-C5GBexperimentdidweseeimprovementintheaggregatedprocessingtime.Thisresultispositivebecausethepower-awarequeryoptimizershowsevengreaterpotentialinpowersavingthanweexpected.Itsdirectoutcomeissubstantialenergysavingsof19%and15.3%. Withthedatabasesizeincreases,morepowerisconsumedforallthreesystems.Thisisunderstandablebecausemoredatawillhavetobereadandprocessedtoanswerthesamequery.Buerhitratewillalsodecreasewhenthedatabaseislarger.Infact,withthe10GBdatabase,systemcontentionbecomesveryhigh{bothCPUanddiskutilizationarecloseto100%formostofthetimeandpowerconsumptiongoesbeyond40w.TheincreaseofdatabasesizehasagreaterimpactonpowerconsumptionoftheoriginalDBMS:powerusagechangesfrom23.8wto41.5winTPC-Handfrom37.5wto42.5winTPC-C.Forthetwopower-awareDBMSs,the35

PAGE 43

DB Time Energy Power Energy Size (W) (min) (kJ) Saving Saving TPC-Hworkloads 0.5GB 21:05 30:06 { { 1 21:1 22:55 28:55 11% 5.1% 0 20:4 23:88 29:23 14% 2:8% 1GB 46:55 106:7 { { 1 31:9 52:05 99:6 13% 6:7% 0 30:9 55:29 102:4 16% 4:1% 10GB 211 525:4 { { 1 34:9 242 508:1 16% 3:3% 0 32:4 263 511:3 22% 3:3% TPC-Cworkloads 1GB 45:34 102:1 { { 1 32:9 48:53 95:79 16% 6:1% 0 31:1 52:43 97:84 17% 4:2% 5GB 123 301:4 { { 1 34:2 120 245:5 16% 19% 0 32:1 132 254:3 20% 15:3% 10GB 223 568:4 { { 1 35:9 243 523:8 16% 8% 0 33:8 263 533:3 21% 7:7% 36

PAGE 44

Wefeedthedatabaseenginewithworkloadscontaining50copiesofonesinglequeryex-tractedfromtheTPC-Hworkload.Aproblemforrunningsingle-queryworkloadsisthatitmaygiveunrealisticqueryprocessingtimeduetoexcessivelyhighcachehitrate.Tosolvethisproblem,weduplicatethedatabaseintovecopiesandeachqueryisrunonadierentcopyofthedatabase.Bythis,wecanavoidtheeectsofdatacachinginourexperiments.Thereare22queriesintheTPC-Htoolswithvariousfeatures.Amongthesequeries,wepicked19becausetheotherthreeareeithersingle-tablequerieswithonesingledominantplanorthosewithextremelylongrunningtimeinourcurrenttestenvironment.Thedatabasesizeis1GB.WecomparethepowerandperformancedatafromrunningtheTPC-Hworkloadwiththecor-respondingestimationsgivenbythe(modied)queryoptimizer,forthepurposeoftestingtheaccuracyofsuchestimations. Letusrststudytheestimatedpowerandperformanceoftheplansselectedbythequeryoptimizer.FromFig.5.3a,wecantellthat11outofthe19queriesshowthepotentialofpowersaving.Theyare:queries1,3,5,6,7,8,11,13,15,17,and18.Forthesequeries,naturally,theexpectedpowersavingcomeswiththecostofelongatedprocessingtime(Fig.5.3b).However,evenwiththeincreaseoftime,mostqueriesstillbearasmallenergyomitting(Fig.5.3c).Astothosequeriesthatarenotexpectedtoprovidepowersavingopportunities,mostofthemaresimplequeries(e.g.,withjoinsofonly2-3tables)thatdonothavealargesearchspace(ofqueryplans)forthequeryoptimizertochoosefrom.Thus,thequeryoptimizerendedupchoosingthesameplanorplanswithverysimilarcosts,nomatterwhichnvalueweuse.For37

PAGE 45

Incomparison,theresultsoftheexperimentsareinteresting.Fig.5.4ashowsthat,forallthequeries,runningthemodiedDBMSengineleadstopowersavingstosomeextent.First,forthecategoryIqueries,(whichshouldnotshowanypowersaving)therearetwocases:1)thosewhoseexecutionplansarenotchangedatall.Theirpowersavingsaretheresultsofrandomsystemerrorsatrun-time;2)thosewhoseexecutionplansaredierentunderdierentnvalues.Thereasonwhytherunningpowerislowerinthesituationsofn=0andn=1couldbe(run-time)systemmodelingerrorsplusmodelingerrorsintheestimationofenergyandperformancecosts.Ontheotherhand,powersavings,althoughnotassignicantasthoseindicatedbytheestimatedvalues(Fig.5.3a),canbeobserved.TherealproblemistheperformancedatashowninFig.5.4b:whenchoosingpower-onlyoptimizer(e.g.,n=0),allquerieshaveunexpectedlylongrunningtime,whichgivesnegativeenergysavingsinmostcases(Fig.5.4c).Thisshowsthatthepower-onlyqueryoptimizercouldbeanundesirablechoiceinthePDBMSdesign.BycomparingtheresultswiththeestimatedcostsinFig.5.3,wealsoseethatthemethodofusingmultiplecopiesofthesamequeryhasseriouslimitationsincapturingthereal(powerandtime)costsofaquery. Whilethedesignofexperimentstotestthecostsofsinglequeryisaninterestingchallengeforfuturework,wecansafelyconcludefromFig.5.4thatpowersavingsdoexist.Salientexamplesarequeries5,15,and17,whichshowsavingsupto25%,whileafewothersshowingsavingsupto20%.Anotherlessonlearnedis:choosingextremevaluesofncouldyieldperformancethatistoopoortobringanydirectenergybenet,itissuggestedtochoosemoderatenvaluesinordertogetbalancedperformanceandenergy.38

PAGE 46

Fig.5.5showsthepowerandperformancedatanormalizedbythethoseoftheoriginalDBMS.InFig.5.5a,weobservepowersavingsinall20cases.However,thesavingsstayinarelativelysteadystateasthepercentageofcategoryIqueryincreases.Ontheotherhand,relativequeryprocessingtimehaslittledierenceinallworkloads(Fig.5.5b).Thus,wecanstillseesavingsontotalenergyconsumption(Fig.5.5c)inmostofthecases.Ascomparedtothepower-only(n=0)optimizer,theonewiththemorebalanced(n=1)costmodelachievedsimilarpowersavingsyetmuchbetterperformance,resultinginmoreenergysavings. WiththepercentageofcategoryIqueriesincreases,powerusagedecreasesinroughlyalinearmannerinallthreesystems(Fig.5.6a).However,theprocessingtimeincreasesexponentially(Fig.5.6b){itincreasesfromabout45minutesinthe1%casetoabout22daysinthe95%case!Naturally,thistrendiscarriedovertotheenergyconsumptiondata(Fig.5.6c),althoughthepowerdecreases. Fromtheseresults,onecanclearlyseethatquerycompositionintheworkloadsdoesnotaectthepowersavingpotentialofourqueryoptimizerdesign.Weachieveaconsiderablemarginofpowersavingdespitethedramaticchangesofqueryprocessingtime.ByvisitingFig.5.5a,itseems20%isareasonableupperboundforpowersavingsonecanreachinanyTPC-Hworkloads.5.4Discussions

PAGE 47

First,evenwiththemarginalenergysavings,theeconomicalimpactoftheenvisionedPDBMSwillbegreat.Forexample,runningasinglehigh-performance300Wserverforoneyearcouldconsume2628KWhofenergy,withanadditional748KWhincoolingthisserver[8].Thetotalenergycostforthissingleserverwouldbe338USdollarsayear(for0.13/KWh)withoutcountingthecostsofairconditioningandpowerdeliverysubsystems[8].Withoursolution,wecansave17-35USdollarsperservereveryyear(considering50%oftotalpowerneededforidlingandthe11-22%powersavingswecanachieve).Again,thisdoesnotcountthecostsavingsachievedbyloweredhardwarefailureandcoolingdemands. Second,weonlyusedaprimitivepower-awareDBMStocapturepower-savingopportunities.Giventhemodelswehaveandthetradeosweidentify,wehavereasonstobelievethatsystempowerusagecouldbefurtherloweredwithmoresophisticatedcostmodelingmechanismsandcontrolalgorithms.Forexample,anadaptivecontrolsolutioncanbedesignedtoguaranteeperformancebychangingtheparametersandeventhestructureofthequeryevaluationmodel.AnotherparadigmwecouldexploreintheforeseeablefutureisPDBMSrunningonserversequippedwithpower-awarehardware,whosepowermodescanbecontrolledbytheDBMSbydesigningtunnelsthatdeliversmessagefromDBMStolowerhardwareviatheOSinterfaces.DuetothereasonsdescribedinSection1,ourmethodcanbecombinedwithOS-levelmethodsforsubstantiallyhigherlevelofpowersavings.40

PAGE 52

Sincethisthesisservesasatestimonyoftheeconomicalandtechnicalvalueofthetopicofpower-awaredatamanagement,abundantfutureresearchopportunitiescanbeforeseen.Ourvisionistobuildanuniedpower-controlframeworkinparalleltotheregularqueryprocessingmodulesintheDBMS.Modelingerrorsinbothenergyandperformanceestimationwouldbeamajorproblemtoattack.Specictasksincluderenedenergymodelsthatcapturethedynamicsinthesystem(insteadofthestaticmodelweusedinthisthesis);searchalgorithmsinthequeryplanspace;advancedcostevaluationcriteria;andmoreimportantly,resourcemanagementalgorithmstolowerthebaselinepowerconsumptionintheserver.45

PAGE 53

\SPECpower ssj2008,http://www.spec.org/power ssj2008,"2008.[Online].Available:http://www.iee.ucsb.edu/greenscale/[2] R.Agrawal,A.Ailamaki,P.A.Bernstein,E.A.Brewer,M.J.Carey,S.Chaudhuri,A.Doan,D.Florescu,M.J.Franklin,H.Garcia-Molina,J.Gehrke,L.Gruenwald,L.M.Haas,A.Y.Halevy,J.M.Hellerstein,Y.E.Ioannidis,H.F.Korth,D.Kossmann,S.Mad-den,R.Magoulas,B.C.Ooi,T.O'Reilly,R.Ramakrishnan,S.Sarawagi,M.Stonebraker,A.S.Szalay,andG.Weikum,\Theclaremontreportondatabaseresearch,"Commun.ACM,vol.52,no.6,pp.56{65,2009.[3] R.AlonsoandS.Ganguly,\EnergyEcientQueryOptimization,"MatsushitaInfoTechLab,Tech.Rep.,1992.[4] D.Anderson,J.Dykes,andE.Riedel,\Morethananinterface|SCSIvs.ATA,"inPro-ceedingsofthe2ndUSENIXConferenceonFileandStorageTechnologies(FAST),2003.[5] M.M.Astrahan,M.W.Blasgen,D.D.Chamberlin,K.P.Eswaran,J.N.Gray,P.P.Griths,W.F.King,R.A.Lorie,P.R.McJones,J.W.Mehl,G.R.Putzolu,I.L.Traiger,B.W.Wade,andV.Watson,\Systemr:relationalapproachtodatabasemanagement,"ACMTrans.DatabaseSyst.,vol.1,no.2,pp.97{137,1976.[6] C.W.Bachman,\Datastructurediagrams,"SIGMISDatabase,vol.1,no.2,pp.4{10,1969.[7] L.BeniniandG.d.Micheli,\System-levelpoweroptimization:techniquesandtools,"ACMTrans.Des.Autom.Electron.Syst.,vol.5,no.2,pp.115{192,2000.46

PAGE 54

R.BianchiniandR.Rajamony,\Powerandenergymanagementforserversystems,"IEEEComputer,vol.37,no.11,pp.68{74,2004.[9] P.Bohrer,E.Elnozahy,T.Keller,M.Kistler,C.Lefurgy,C.McDowell,andR.Rajamony,\Thecaseforpowermanagementinwebservers,"inPower-AwareComputing,R.GraybillandR.Melhem,Eds.KlewerAcademic/PlenumPublishers,2002,pp.261{287.[10] D.BrooksandM.Martonosi,\Dynamicthermalmanagementforhigh-performancemi-croprocessors."inProceedingsofthe7thInternationalSymposiumonHigh-PerformanceComputerArchitecture(HPCA),2001.[11] D.Brooks,P.Bose,S.Schuster,H.Jacobson,P.Kudva,A.Buyuktosunoglu,J.Wellman,V.Zyuban,M.Gupta,andP.Cook,\Power-awaremicroarchitecture:designandmodelingchallengesfornext-generationmicroprocessors,"Micro,IEEE,vol.20,no.6,pp.26{44,Nov/Dec2000.[12] J.S.Chase,D.C.Anderson,P.N.Thakar,A.M.Vahdat,andR.P.Doyle,\Managingenergyandserverresourcesinhostingcenters,"inProceedingsofthe18thACMSymposiumonOperatingSystemsPrinciples(SOSP),2001.[13] Y.Chen,T.Wang,J.M.Hellserstein,andR.H.Katz,\EnergyEciencyofMapReduce,http://www.eecs.berkeley.edu/Research/Projects/Data/105613.html,"2008.[14] Y.Chen,A.Das,W.Qin,A.Sivasubramaniam,Q.Wang,andN.Gautam,\Managingserverenergyandoperationalcostsinhostingcenters,"inProceedingsofACMInterna-tionalConferenceonMeasurementandModelingofComputerSystems(SIGMETRICS),2005.[15] A.P.Conversion,\DeterminingTotalCostofOwnershipforDataCentersandNetworkRoomInfrastructure.ftp://www.apcmedia.com/salestools/CMRP-5T9PQG R2 EN.pdf,"2005.47

PAGE 55

R.P.Dick,G.Lakshminarayana,A.Raghunathan,andN.K.Jha,\Poweranalysisofembeddedoperatingsystems,"inDAC'00:Proceedingsofthe37thAnnualDesignAu-tomationConference.NewYork,NY,USA:ACM,2000,pp.312{315.[17] M.Elnozahy,M.Kistler,andR.Rajamony,\Energy-ecientserverclusters,"inProceed-ingsofthe2ndWorkshoponPower-AwareComputingSystems,2002.[18] X.Fan,W.-D.Weber,andL.A.Barroso,\Powerprovisioningforawarehouse-sizedcom-puter,"inISCA'07:Proceedingsofthe34thannualinternationalsymposiumonComputerarchitecture.NewYork,NY,USA:ACM,2007,pp.13{23.[19] S.G.W.C.H.D.A.W.M.B.T.C.A.F.R.E.G.FayChang,JereyDean,\Bigtable:Adistributedstoragesystemforstructureddata,"Proceedingsofthe7thUSENIXSymposiumonOperatingSystems,pp.205{218,2006.[20] W.Felter,K.Rajamani,T.Keller,andC.Rusu,\Aperformance-conservingapproachforreducingpeakpowerconsumptioninserversystems,"inProceedingsofthe19thannualinternationalconferenceonSupercomputing(ICS),2005.[21] S.Furber,A.Efthymiou,J.Garside,D.Lloyd,M.Lewis,andS.Temple,\Powermanage-mentintheamuletmicroprocessors,"DesignandTestofComputers,IEEE,vol.18,no.2,pp.42{52,2001.[22] H.Garcia-Molina,J.D.Ullman,andJ.Widom,DatabaseSystems:TheCompleteBook.UpperSaddleRiver,NJ,USA:PrenticeHallPress,2008.[23] R.GonzalezandM.Horowitz,\Energydissipationingeneralpurposeprocessors,"inLowPowerElectronics,1995.,IEEESymposiumon,Oct1995,pp.12{13.[24] T.Heath,A.P.Centeno,P.George,L.Ramos,Y.Jaluria,andR.Bianchini,\MercuryandFreon:Temperatureemulationandmanagementforserversystems,"inProceedingsofthe12thInternationalConferenceonArchitecturalSupportforProgrammingLanguagesandOperatingSystems(ASPLOS),2006.48

PAGE 56

J.HenkelandY.Li,\Avalanche:anenvironmentfordesignspaceexplorationandopti-mizationoflow-powerembeddedsystems,"VeryLargeScaleIntegration(VLSI)Systems,IEEETransactionson,vol.10,no.4,pp.454{468,Aug2002.[26] T.Horvath,T.Abdelzaher,K.Skadron,andX.Liu,\Dynamicvoltagescalinginmulti-tierwebserverswithend-to-enddelaycontrol,"IEEETransactionsonComputers,vol.56,no.4,pp.444{458,2007.[27] H.Huang,C.Lefurgy,K.Rajamani,T.Keller,E.vanHensbergen,F.Rawson,andK.G.Shin,\CooperativeSoftware-HardwarePowerManagementforMainMemory,"inPower-AwareComputerSystems-4thInternationalWorkshop(PACS),B.FalsaandT.N.Vijaykumar,Eds.Springer,December2004.[28] IDC,\SolutionsforDataCenters'ThermalChallenges,http://www.blade.org/docs/wp/idc cool blue whitepaper.pdf,"January2007.[Online].Available:www.blade.org/docs/wp/idc cool blue whitepaper.pdf[29] InstituteforEnergyEciency,\UCSantaBarbara,http://www.iee.ucsb.edu/greenscale/,"2008.[Online].Available:http://www.iee.ucsb.edu/greenscale/[30] A.IyerandD.Marculescu,\Powerawaremicroarchitectureresourcescaling,"inDATE'01:ProceedingsoftheconferenceonDesign,automationandtestinEurope.Piscataway,NJ,USA:IEEEPress,2001,pp.190{196.[31] M.J.Kamfonas,\Recursivehierarchies:Therelationaltaboo!"TheRelationalJournal-October/November,1992.[32] J.G.Koomey,\EstimatingTotalPowerConsumptionbyServersintheU.S.andtheWorld,"LawrenceBerkeleyNationalLaboratory,URL:http://hightech.lbl.gov/documents/DATA CENTERS/svrpwrusecompletenal.pdf,Tech.Rep.,February2007.[Online].Available:http://hightech.lbl.gov/documents/DATA CENTERS/svrpwrusecompletenal.pdf49

PAGE 57

U.H.LABarroso,\Thecaseforenergy-proportionalcomputing,"IEEECOMPUTERSOCIETY,IEEE,2007.[34] J.Laudon,\Performance/watt:thenewserverfocus,"SIGARCHComputerArchitectureNews,vol.33,no.4,pp.5{13,2005.[35] C.Lefurgy,K.Rajamani,F.Rawson,W.Felter,M.Kistler,andT.W.Keller,\Energymanagementforcommercialservers,"IEEEComputer,vol.36,no.12,pp.39{48,2003.[36] Y.LiandJ.Henkel,\Aframeworkforestimationandminimizingenergydissipationofembeddedhw/swsystems,"inDAC'98:Proceedingsofthe35thannualDesignAutomationConference.NewYork,NY,USA:ACM,1998,pp.188{193.[37] W.Liao,L.He,andK.Lepak,\Temperatureandsupplyvoltageawareperformanceandpowermodelingatmicroarchitecturelevel,"Computer-AidedDesignofIntegratedCircuitsandSystems,IEEETransactionson,vol.24,no.7,pp.1042{1053,July2005.[38] J.Liu,P.H.Chou,N.Bagherzadeh,andF.Kurdahi,\Aconstraint-basedapplicationmodelandschedulingtechniquesforpower-awaresystems,"inCODES'01:ProceedingsoftheninthinternationalsymposiumonHardware/softwarecodesign.NewYork,NY,USA:ACM,2001,pp.153{158.[39] ||,\Power-awareschedulingundertimingconstraintsformission-criticalembeddedsys-tems,"inDAC'01:Proceedingsofthe38thannualDesignAutomationConference.NewYork,NY,USA:ACM,2001,pp.840{845.[40] J.LorchandA.Smith,\Softwarestrategiesforportablecomputerenergymanagement,"PersonalCommunications,IEEE,vol.5,no.3,pp.60{73,Jun1998.[41] Y.-H.Lu,L.Benini,andG.D.Micheli,\Operating-systemdirectedpowerreduction,"inProceedingsoftheInternationalSymposiumonLowPowerElectronicsandDesign(ISLPED),2000.50

PAGE 58

Y.-H.Lu,T.Simunic,andG.DeMicheli,\Softwarecontrolledpowermanagement,"inCODES'99:ProceedingsoftheseventhinternationalworkshoponHardware/softwarecodesign.NewYork,NY,USA:ACM,1999,pp.157{161.[43] J.Manko,R.Kravets,andE.Blevis,\SomeComputerScienceIssuesinCreatingaSustainableWorld,"IEEEComputer,vol.41,no.8,pp.102{105,August2008.[44] A.Martin,M.Nystrom,andP.Penzes,\Et2:AMetricforTimeandEnergyEciencyofComputation,"inPower-AwareComputing,R.GraybillandR.Melhem,Eds.KlewerAcademic/PlenumPublishers,2002,pp.293{313.[45] T.Mudge,\Power:AFirst-ClassArchitecturalDesignConstraint,"IEEEComputer,vol.34,no.3,pp.52{58,April2001.[46] R.Nambiar,\3-YearEnergyCostinTPCBenchmarks,"2008.[Online].Available:http://library.hp.com/techcon07/75.pdf[47] C.Patel,C.Bash,R.Sharma,M.Beitelmal,andR.Friedrich,\Smartcoolingofdatacenters,"inProceedingsoftheASMEInterpackConference,2003.[48] M.PoessandR.O.Nambiar,\Energycost,thekeychallengeoftoday'sdatacenters:apowerconsumptionanalysisofTPC-Cresults,"ProceedingsoftheVeryLargeDataBases(VLDB)Endowment,vol.1,no.2,pp.1229{1240,2008.[49] L.RamosandR.Bianchini,\C-Oracle:Predictivethermalmanagementfordatacenters,"inProceedingsofthe14thIEEEInternationalSymposiumonHigh-PerformanceComputerArchitecture(HPCA),2008.[50] S.Rivoire,M.A.Shah,P.Ranganathan,andC.Kozyrakis,\JouleSort:abalancedenergy-eciencybenchmark,"inProceedingsoftheACMInternationalConferenceonManage-mentofData(SIGMOD),2007,pp.365{376.[51] V.Sharma,A.Thomas,T.Abdelzaher,K.Skadron,andZ.Lu,\Power-awareQoSman-agementinwebservers,"inProceedingsofthe24thIEEEReal-TimeSystemsSymposium(RTSS),2003.51

PAGE 59

K.Skadron,M.R.Stan,K.Sankaranarayanan,W.Huang,S.Velusamy,andD.Tar-jan,\Temperature-awaremicroarchitecture:Modelingandimplementation,"ACMTrans.Archit.CodeOptim.,vol.1,no.1,pp.94{125,2004.[53] M.Stonebraker,D.J.Abadi,A.Batkin,X.Chen,M.Cherniack,M.Ferreira,E.Lau,A.Lin,S.Madden,E.O'Neil,P.O'Neil,A.Rasin,N.Tran,andS.Zdonik,\C-store:acolumn-orienteddbms,"inVLDB'05:Proceedingsofthe31stinternationalconferenceonVerylargedatabases.VLDBEndowment,2005,pp.553{564.[54] UnitedStatesEnvironmentalProtectionAgency,\Reporttocongressonserveranddatacenterenergyeciency,"http://www.energystar.gov/ia/partners/proddevelopment/downloads/EPA Datacenter Report Congress Final1.pdf,Tech.Rep.,February2007.[Online].Available:http://hightech.lbl.gov/documents/DATA CENTERS/svrpwrusecompletenal.pdf[55] X.WangandM.Chen,\Cluster-levelfeedbackpowercontrolforperformanceoptimiza-tion,"inProceedingsofthe14thIEEEInternationalSymposiumonHigh-PerformanceComputerArchitecture(HPCA),2008.[56] M.Weiser,B.Welch,A.J.Demers,andS.Shenker,\SchedulingforreducedCPUenergy,"inProceedingsoftheFirstSymposiumonOperatingSystemsDesignandImplementation(OSDI),1994.[57] C.Xian,Y.-H.Lu,andZ.Li,\AProgrammingEnvironmentwithRuntimeEnergyCharac-terizationforEnergy-AwareApplications,"inProceedingsoftheInternationalSymposiumonLowPowerElectronicsandDesign(ISLPED),2007,pp.141{146.[58] L.-T.YehandR.C.Chu,ThermalManagementofMicroelectronicEquipment:HeatTrans-ferTheory,AnalysisMethods,andDesignPractices.ASMEPress,2002.[59] H.Zeng,C.S.Ellis,A.R.Lebeck,andA.Vahdat,\ECOSystem:managingenergyasarstclassoperatingsystemresource,"inProceedingsofthe10thInternationalConferenceonArchitecturalSupportforProgrammingLanguagesandOperatingSystems(ASPLOS),2002.52

PAGE 60

Y.ZhuandF.Mueller,\FeedbackEDFschedulingexploitingdynamicvoltagescaling,"inProceedingsofthe10thIEEEReal-TimeTechnologyandApplicationsSymposium(RTAS),2004.53