ARGONNENATIONALLABORATORY 9700SouthCassAvenue Argonne,Illinois60439. BenchmarkingOptimizationSoftwarewithPerformanceProles



Similar documents
An Overview Of Software For Convex Optimization. Brian Borchers Department of Mathematics New Mexico Tech Socorro, NM

Conic optimization: examples and software

Solving NP Hard problems in practice lessons from Computer Vision and Computational Biology

Modern Optimization Methods for Big Data Problems MATH11146 The University of Edinburgh

Finding Liveness Errors with ACO

Uncertainty modeling revisited: What if you don t know the probability distribution?

Discrete Optimization

Dantzig-Wolfe bound and Dantzig-Wolfe cookbook

Optimal Hiring of Cloud Servers A. Stephen McGough, Isi Mitrani. EPEW 2014, Florence

Performance of Optimization Software - an Update

A CP Scheduler for High-Performance Computers

Bandwidth Allocation in a Network Virtualization Environment

GAMS, Condor and the Grid: Solving Hard Optimization Models in Parallel. Michael C. Ferris University of Wisconsin

Illustration (and the use of HLM)

Portfolio Distribution Modelling and Computation. Harry Zheng Department of Mathematics Imperial College

Gate Delay Model. Estimating Delays. Effort Delay. Gate Delay. Computing Logical Effort. Logical Effort


TOMLAB - For fast and robust largescale optimization in MATLAB

Lecture 3. Linear Programming. 3B1B Optimization Michaelmas 2015 A. Zisserman. Extreme solutions. Simplex method. Interior point method

Warshall s Algorithm: Transitive Closure

Lecture 11: 0-1 Quadratic Program and Lower Bounds

Minimizing the Number of Machines in a Unit-Time Scheduling Problem

Algorithm Design and Analysis

Numerical methods for American options

Error Log Processing for Accurate Failure Prediction. Humboldt-Universität zu Berlin

Discuss the size of the instance for the minimum spanning tree problem.

A numerically adaptive implementation of the simplex method

Airlift: Video Conferencing as a Cloud Service using Inter- Datacenter Networks

A Tool to aid nmon analysis and utility


Matpower 4.1 User s Manual

Data Mining on Streams

Hardware/Software Codesign

Optimal Income Taxation with Multidimensional Types

Toward Variability Management to Tailor High Dimensional Index Implementations

FUZZY CLUSTERING ANALYSIS OF DATA MINING: APPLICATION TO AN ACCIDENT MINING SYSTEM

Spare part inventory control for an aircraft component repair shop

HYBRID GENETIC ALGORITHMS FOR SCHEDULING ADVERTISEMENTS ON A WEB PAGE

INTERIOR-POINT METHODS FOR NONCONVEX NONLINEAR PROGRAMMING: FILTER METHODS AND MERIT FUNCTIONS

DATA VERIFICATION IN ETL PROCESSES

Extension of measure

! Solve problem to optimality. ! Solve problem in poly-time. ! Solve arbitrary instances of the problem. !-approximation algorithm.

BUSINESS ANALYTICS. Data Pre-processing. Lecture 3. Information Systems and Machine Learning Lab. University of Hildesheim.

Linear Programming Sensitivity Analysis

Parallelization Strategies for Multicore Data Analysis

THE PROBLEM WORMS (1) WORMS (2) THE PROBLEM OF WORM PROPAGATION/PREVENTION THE MINIMUM VERTEX COVER PROBLEM

Copula model estimation and test of inventory portfolio pledge rate

Dušan Bernát

Lab 5 Linear Regression with Within-subject Correlation. Goals: Data: Use the pig data which is in wide format:

11. APPROXIMATION ALGORITHMS

Preprint Ayalew Getachew Mersha, Stephan Dempe Feasible Direction Method for Bilevel Programming Problem ISSN

An Additive Neumann-Neumann Method for Mortar Finite Element for 4th Order Problems

Mixed Integer Linear Programming in R

Lecture 5: Logical Effort

Optimization of IP Load-Balanced Routing for Hose Model

SBB: A New Solver for Mixed Integer Nonlinear Programming

Probabilistic Models for Big Data. Alex Davies and Roger Frigola University of Cambridge 13th February 2014

Load Balancing by MPLS in Differentiated Services Networks

IEOR 4404 Homework #2 Intro OR: Deterministic Models February 14, 2011 Prof. Jay Sethuraman Page 1 of 5. Homework #2

A Service Design Problem for a Railway Network

ANALYSIS, THEORY AND DESIGN OF LOGISTIC REGRESSION CLASSIFIERS USED FOR VERY LARGE SCALE DATA MINING

Understanding the Impact of Weights Constraints in Portfolio Theory

GAMS Productivity - Performance - Reliability

PuLP: A Linear Programming Toolkit for Python

Medication Prescribing Practices of CRNPs in Mental Health Programs and Mental Retardation Programs BY:

BEFORE THE INSURANCE COMMISSIONER OF THE COMMONWEALTH OF PENNSYLVANIA DECISION AND ORDER

5 INTEGER LINEAR PROGRAMMING (ILP) E. Amaldi Fondamenti di R.O. Politecnico di Milano 1

Enterprise Applications in the Cloud: Non-virtualized Deployment

Windows Scheduled Task and PowerShell Scheduled Job Management Pack Guide for Operations Manager 2012

Lemma 5.2. Let S be a set. (1) Let f and g be two permutations of S. Then the composition of f and g is a permutation of S.

MARKETING TIP. Ask to share with a friend

HURDLE AND SELECTION MODELS Jeff Wooldridge Michigan State University BGSE/IZA Course in Microeconometrics July 2009

Fundamentals and Recent Developments of the GAMS System

SOLVING EQUATIONS WITH EXCEL

Airport Planning and Design. Excel Solver

Modeling and Performance Evaluation of Computer Systems Security Operation 1

Transcription:

ARGONNENATIONALLABORATORY 9700SouthCassAvenue Argonne,Illinois60439 BenchmarkingOptimizationSoftwarewithPerformanceProles MathematicsandComputerScienceDivision ElizabethD.DolanandJorgeJ.More PreprintANL/MCS-P861-1200 January2001 Energy,underContractW-31-109-Eng-38,andbytheNationalScienceFoundation(ChallengesinComputationalScience)grantCDA-9726385and(InformationTechnologyResearch)grantCCR-0082807. ThisworkwassupportedbytheMathematical,Information,andComputationalSciences DivisionsubprogramoftheOceofAdvancedScienticComputing,U.S.Departmentof

Contents 4CaseStudy:OptimalControlandParameterEstimationProblems 2PerformanceEvaluation 3BenchmarkingData 1Introduction 6341 5CaseStudy:TheFullCOPS 6CaseStudy:LinearProgramming 7Conclusions Acknowledgments 10 11 9 References 12

BenchmarkingOptimizationSoftwarewithPerformanceProles ElizabethD.DolanyandJorgeJ.Morez 1Introduction Weproposeperformanceproles distributionfunctionsforaperformancemetric asatoolforbenchmarkingandcomparingoptimizationsoftware.weshowthatperformanceprolescombinethebestfeaturesofothertoolsforperformanceevaluation. Abstract withtheevaluationandperformanceofoptimizationcodes.asrecentexamples,wecite [1,2,3,4,6,12,17]. Mittelmann'seortshavegainedthemostnotice,otherresearchershavebeenconcerned Thebenchmarkingofoptimizationsoftwarehasrecentlygainedconsiderablevisibility.Hans themaintechnicalissuesaddressedinthispaper.mostbenchmarkingeortsinvolvetables Mittlemann's[13]workonavarietyofoptimizationsoftwarehasfrequentlyuncovered displayingtheperformanceofeachsolveroneachproblemforasetofmetricssuchascpu decienciesinthesoftwareandhasgenerallyledtosoftwareimprovements.although time,numberoffunctionevaluations,oriterationcountsforalgorithmswhereaniteration impliesacomparableamountofwork.failuretodisplaysuchtablesforasmalltestset wouldbeagrossomission,buttheytendtobeoverwhelmingforlargetestsets.inallcases, theinterpretationoftheresultsfromthesetablesisoftenasourceofdisagreement. Theinterpretationandanalysisofthedatageneratedbythebenchmarkingprocessare researcherstotryvarioustoolsforanalyzingthedata.thesolver'saverageorcumulative mance[1,4,6].asaresult,asmallnumberofthemostdicultproblemscantendto dominatetheseresults,andresearchersmusttakepainstogiveadditionalinformation.anotherdrawbackisthatcomputingaveragesortotalsforaperformancemetricnecessitates discardingproblemsforwhichanysolverfailed,eectivelybiasingtheresultsagainstthe totalforeachperformancemetricoverallproblemsissometimesusedtoevaluateperfor- Thequantitiesofdatathatresultfrombenchmarkingwithlargetestsetshavespurred subprogramoftheoceofadvancedscienticcomputing,u.s.departmentofenergy,undercontract W-31-109-Eng-38,andbytheNationalScienceFoundation(ChallengesinComputationalScience)grant valuecanbeassignedforfailedsolverattempts,butthisrequiresasubjectivechoicefor mostrobustsolvers.asanalternativetodisregardingsomeoftheproblems,apenalty separatetable. thepenalty.mostresearcherschoosetoreportthenumberoffailuresonlyinafootnoteor (more@mcs.anl.gov). ComputerScienceDivision,ArgonneNationalLaboratory,Argonne,Illinois60439(dolan@mcs.anl.gov). CDA-9726385and(InformationTechnologyResearch)grantCCR-0082807. ThisworkwassupportedbytheMathematical,Information,andComputationalSciencesDivision zmathematicsandcomputersciencedivision,argonnenationallaboratory,argonne,illinois60439 ydepartmentofelectricalandcomputerengineering,northwesternuniversity,andmathematicsand 1

dierencebetweensolvertimes[4])appearstobeaviablewayofensuringthataminority place,usuallyfork=1;2;3.rankingthesolvers'performanceforeachproblemhelps [4,6,15,17].Inotherwords,theycountthenumberoftimesthatasolvercomesinkth preventaminorityoftheproblemsfromundulyinuencingtheresults.informationonthe sizeoftheimprovement,however,islost. Toaddresstheshortcomingsofthepreviousapproach,someresearchersrankthesolvers oftheproblemsdonotdominatetheresults,butinourtestingwehavewitnessedlargeleaps inquartilevaluesofaperformancemetric,ratherthangradualtrends.ifonlyquartiledata isused,theninformationontrendsoccurringbetweenonequartileandthenextislost;and wemustassumethatthejourneyfromonepointtoanotherproceedsatamoderatepace. Also,inthespeciccaseofcontrastingthedierencesbetweensolvertimes,thecomparison Comparingthemediansandquartilesofsomeperformancemetric(forexample,the isthatifresultsaremixed,interpretingquartiledatamaybenoeasierthanusingtheraw data;anddealingwithcomparisonsofmorethantwosolversmightbecomeunwieldy. failstoprovideanyinformationontherelativesizeoftheimprovement.analdrawback denethebordersofverycompetitiveandcompetitive. appearsin[2],withsolversratedbythepercentageofproblemsforwhichasolver'stimeis termedverycompetitiveorcompetitive.theratioapproachavoidsmostofthediculties thatwehavediscussed,providinginformationonthepercentimprovementandeliminating thenegativeeectsofallowingasmallportionoftheproblemstodominatetheconclusions. Themaindisadvantageofthisapproachliesintheauthor'sarbitrarychoiceoflimitsto Theideaofcomparingsolversbytheratioofonesolver'sruntimetothebestruntime mulative)distributionfunctionforaperformancemetric.inthispaperweusetheratioof thecomputingtimeofthesolverversusthebesttimeofallofthesolversastheperformance metric.section3providesananalysisofthetestsetandsolversusedinthebenchmark theperformanceofoptimizationsoftware.theperformanceproleforasolveristhe(cu- resultsofsections4and5.thisanalysisisnecessarytounderstandthelimitationsofthe benchmarkingprocess. InSection2,weintroduceperformanceprolesasatoolforevaluatingandcomparing parameterchoicesandtheneedtodiscardsolverfailuresfromtheperformancedata. inuenceofasmallnumberofproblemsonthebenchmarkingprocessandthesensitivityof resultsassociatedwiththerankingofsolvers.performanceprolesprovideameansofvisualizingtheexpectedperformancedierenceamongmanysolvers,whileavoidingarbitrary withversion2.0ofthecops[7]testset.weshowthatperformanceproleseliminatethe WeconcludeinSection6byshowinghowperformanceprolesapplytothedata[13] Sections4and5demonstratetheuseofperformanceproleswithresults[8]obtained theuseofperformanceprolesandalsoshowsthatperformanceprolescanbeappliedto ofmittelmannforlinearprogrammingsolvers.thissectionprovidesanothercasestudyof awiderangeofperformancedata. 2

informationofinterestsuchasthenumberoffunctionevaluationsandthecomputingtime. Inthissectionweintroducethenotionofaperformanceproleasameanstoevaluateand comparetheperformanceofthesolversonatestsetp. BenchmarkresultsaregeneratedbyrunningasolveronasetPofproblemsandrecording 2PerformanceEvaluation measures.foreachproblempandsolvers,wedene putingtimeasaperformancemeasure;although,theideasbelowcanbeusedwithother Weassumethatwehavenssolversandnpproblems.Weareinterestedinusingcom- performanceratio bysolverswiththebestperformancebyanysolveronthisproblem;thatis,weusethe If,forexample,thenumberoffunctionevaluationsistheperformancemeasureofinterest, settp;saccordingly. Werequireabaselineforcomparisons.Wecomparetheperformanceonproblemp tp;s=computingtimerequiredtosolveproblempbysolvers. solversdoesnotsolveproblemp.theperformanceofsolversonanygivenproblemmay solver.ifwedene beofinterest,butwewouldliketoobtainanoverallassessmentoftheperformanceofthe WeassumethataparameterMp;sforallp;sischosen,andp;s=Mifandonlyif ps()=1npsizenp2p:p;so; p;s= minftp;s:1snsg: possibleratio.thefunctionpsisthe(cumulative)distributionfunctionfortheperformance problemsthatarelikelytooccurinapplications,thensolverswithlargeprobabilityps() ratioẇeusethetermperformanceproleforthedistributionfunctionofaperformancemetric.ourclaimisthataplotoftheperformanceprolerevealsallofthemajorperformance characteristics.inparticular,ifthesetofproblemspissuitablylargeandrepresentativeof thenps()istheprobabilitythataperformanceratiop;siswithinafactorofthebest aretobepreferred. versusaproblemparameter.forexample,higham[11,pages296{297]plotstheratio =kak1,whereistheestimateforthel1normofamatrixaproducedbythelapack conditionnumberestimator.notethatinhigham'suseofthetermperformanceprole thereisnoattemptatdeterminingadistributionfunction. function,continuousfromtherightateachbreakpoint.thevalueofps(1)istheprobability Thetermperformanceprolehasalsobeenusedforaplotofsomeperformancemetric numberofwins,weneedonlytocomparethevaluesofps(1)forallofthesolvers. thatthesolverwillwinovertherestofthesolvers.thus,ifweareinterestedonlyinthe Theperformanceproleps:R7![0;1]forasolverisanondecreasing,piecewiseconstant 3

istheprobabilitythatthesolversolvesaproblem.thus,ifweareinterestedonlyinsolvers resultofthisconvention,ps(m)=1,and thatp;s2[1;m]andthatp;s=monlywhenproblempisnotsolvedbysolvers.asa Thedenitionoftheperformanceproleforlargevaluesrequiressomecare.Weassume withahighprobabilityofsuccess,thenweneedtocomparethevaluesofps( M)forall solversandchoosethesolverswiththelargestvalue.thevalueofps( M)canbereadilyseen for2[s;m)forsomes<m. inaperformanceprolebecausepsatlinesforlargevaluesof;thatis,ps()=ps( M) ps( M)lim! Mps() asourceofdisagreementbecausethereisnoconsensusonhowtochooseproblems.thecops problemsareselectedtobeinterestinganddicult,butthiscriteriaissubjective.because 3BenchmarkingData themainpurposeofourtestingistouncoverthebestgeneralsolveronawiderangeof problems,weincludeproblemswithmultipleminimaandproblemswithalargenumberof ThetimingdatausedtocomputetheperformanceprolesinSections4and5isgenerated degreesoffreedom.foreachoftheapplicationsinthecopssetweusefourinstancesofthe withthecopstestset,whichcurrentlyconsistsofseventeendierentapplications,all applicationobtainedbyvaryingaparameterintheapplication,forexample,thenumber modelsintheampl[9]modelinglanguage.thechoiceofthetestproblemsetpisalways mationapplicationsinthecopsset,whilethediscussioninsection5coversthecomplete ofgridpointsinadiscretization. performanceresults.accordingly,weprovidehereinformationspecictothissubsetofthe COPSproblemsaswellasananalysisofthetestsetasawhole.Table3.1givesthequartiles forthreeproblemparameters:thenumberofvariablesn,thenumberofconstraints,and theratio(n ne)=n,whereneisthenumberofequalityconstraints.intheoptimization literature,n neisoftencalledthedegreesoffreedomoftheproblem,sinceitisanupper Section4focusesononlythesubsetoftheelevenoptimalcontrolandparameteresti- throughoutthetestsetandshowsthatatleastthree-fourthsoftheproblemshavethe boundonthenumberofvariablesthatarefreeatthesolution. AnotherfeatureoftheCOPSsubsetisthattheequalityconstraintsaretheresultofeither intherange[1;50]becauseotherbenchmarkingproblemsetstendtohaveapreponderance numberofvariablesnintheinterval[400;5000].ouraimwastoavoidproblemswherenwas ofproblemswithninthisrange.themaindierencebetweenthefullcopssetandthe COPSsubsetisthattheCOPSsubsetismoreconstrainedwithnen=2foralltheproblems. ThedatainTable3.1isfairlyrepresentativeofthedistributionoftheseparameters settingtheoutputlevelsothatwecangatherthedataweneed,increasingtheiteration dierenceorcollocationapproximationstodierentialequations. Weranournalcompleterunswiththesameoptionsforallmodels.Theoptionsinvolve 4

to5000.noneofthefailureswerecordinthenaltrialsincludeanysolvererrormessages abouthavingviolatedtheselimits. limitsasmuchasallowed,andincreasingthesuper-basicslimitsforminosandsnopt Numberofconstraints0150498159850485140080016014797 Degreesoffreedom(%)0.01.033.2100.0100.00.00.419.833.149.9 Numberofvariables4840010002402500010044989920004815 minq1q2q3maxminq1q2q3max Table3.1:ProblemdataforCOPStestset 023148401500005992011198 FullCOPS COPSsubset soastominimizetheeectofuctuationinthemachineload.thescripttracksthewallclocktimefromthestartofthesolve,killinganyprocessthatruns3,600seconds,which recordingthewall-clocktimeaswellasthecombinationofamplsystemtime(tointerpret themodelandcomputevaryingamountsofderivativeinformationrequiredbyeachsolver) andsolvertimeforeachmodelvariation.werepeatthecycleforanymodelforwhichone wedeclareunsuccessful,andbeginningthenextsolve.wecyclethroughalltheproblems, Thescriptforgeneratingthetimingdatasendsaproblemtoeachsolversuccessively, ofthesolver'sampltimesandthewall-clocktimesdierbymorethan10percent.to furtherensureconsistency,wehaveveriedthattheampltimeresultswepresentcouldbe requirementscanalsoplayanimportantrole.inparticular,solversthatusedirectlinear criteriaarenotlikelytochangecomputingtimesbymorethanafactoroftwo.memory intheamplenvironment.inanycase,dierencesincomputingtimeduetothestopping thesolvers.ideallywewouldhaveusedthesamestoppingcriteria,butthisisnotpossible runningsolaris7. reproducedtowithin10percentaccuracy.allcomputationsweredoneonasparcultra2 equationsolversareoftenmoreecientintermsofcomputingtimeprovidedthereisenough memory. Wehaveignoredtheeectsofthestoppingcriteriaandthememoryrequirementsof mation.theuseofsecond-orderinformationcanreducethenumberofiterations,but useonlyrst-orderinformation,whilelancelotandloqoneedsecond-orderinfor- thecostperiterationusuallyincreases.inaddition,obtainingsecond-orderinformationis morecostlyandmaynotevenbepossible.minosandsnoptarespecicallydesigned LANCELOTandLOQO.Asanexampleofcomparingsolverswithsimilarrequirements, forproblemswithamodestnumberofdegreesoffreedom,whilethisisnotthecasefor Thesolversthatwebenchmarkherehavedierentrequirements.MINOSandSNOPT Section6showstheperformanceoflinearprogrammingsolvers. 5

4CaseStudy:OptimalControlandParameterEstimationProblems WenowexaminetheperformanceprolesofLANCELOT[5],MINOS[14],SNOPT[10], andloqo[16]onthesubsetoftheoptimalcontrolandparameterestimationproblemsin thecops[7]testset.figures4.1and4.2showtheperformanceprolesindierentranges providethenecessaryobjectiveinformationforreasonablesubjectiveanalysisofalargetest set.figure4.1showstheperformanceprolesofthefoursolversforsmallvaluesof. Byshowingtheratiosofsolvertimes,weeliminateanyweightofimportancethattaking toreectvariousareasofinterest.ourpurposeistoshowhowtheperformanceproles ofbeingtheoptimalsolver)andthattheprobabilitythatloqoisthewinneronagiven straighttimedierencesmightgivetotheproblemsthatrequirealongruntimeofevery solversreceivetheirduecreditforcompletingproblemsforwhichoneormoreoftheother solversfails.inparticular,1 ps()isthefractionofproblemsthatthesolvercannotsolve withinafactorofthebestsolver,includingproblemsforwhichthesolverinquestionfails. solver.wendnoneedtoeliminateanytestproblemsfromdiscussion.forthisreason, showsthattheprobabilitythatthesetwosolverscansolveajobwithinafactor4ofthe MINOS,butitsperformancebecomesmuchmorecompetitiveifweextendourofinterest bestsolverisonlyabout70%.snopthasalowernumberofwinsthaneitherloqoor problemisabout:61.ifwechoosebeingwithinafactorof4ofthebestsolverasthescope ofourinterest,theneitherloqoorminoswouldsuce;buttheperformanceprole FromthisgureitisclearthatLOQOhasthemostwins(hasthehighestprobability successfully,thensnoptcapturesourattentionwithitsabilitytosolveover90%ofthis areinterestedinthesolverthatcansolve75%oftheproblemswiththegreatesteciency, thenminosstandsout.ifweholdtomorestringentprobabilitiesofcompletingasolve COPSsubset,asdisplayedbytheheightofitsperformanceprolefor>40.Thisgraph to7ḟigure4.2showstheperformanceprolesforallthesolversintheinterval[1;100].ifwe interval[15;40]. percentageoftheproblems.anotherpointofinterestisthatloqo,minos,andsnopt eachhavethebestprobabilityps()forinsomeinterval,withsimilarperformanceinthe displaysthepotentialforlargediscrepanciesintheperformanceratiosonasubstantial betweenquartilesdoesnotnecessarilyproceedlinearly;hence,wereallyloseinformation forsnoptand13:9forloqo.bylookingatfigures4.1and4.2,weseethatprogress MINOSalsosharerstquartilevaluesof1.Inotherwords,thesetwosolversarethe bestsolversonatleast25%oftheproblems.loqobestsminos'smedianvaluewith1 comparedwith2:4,butminoscomesbackwithathirdquartileratioof4:3versus12:6 valuesoftimeratios.thetopthreesolversshareaminimumratioof1,andloqoand Anobservationthatemergesfromtheseguresisthelackofconsistencyinquartile andnoobviousalternativevalueexists.asanalternativetoprovidingonlyquartilevalues, ifwedonotprovidethefulldata.also,themaximumratiowouldbemforourtesting, 6

1 Performance Profile on Subset of COPS 0.9 0.8 P( ( t p,s / min {t p,s : 1 s n s } ) τ) 0.7 Figure4.1:Performanceproleon[0;10] 0.6 0.5 0.4 LANCELOT MINOS 0.3 SNOPT LOQO 0.2 0.1 0 1 2 3 4 5 6 7 8 9 10 τ 1 Performance Profile on Subset of COPS 0.9 P( ( t p,s / min {t p,s : 1 s n s } ) τ) 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 7 LANCELOT MINOS SNOPT 10 20 30 40 50 60 70 80 90 100 τ Figure4.2:Performanceproleon[0;100]LOQO

however,theperformanceproleyieldsmuchmoreinformationaboutasolver'sstrengths andweaknesses. of2forthescale.inotherwords,weplot fullimplicationsofourtestdataregardingthesolvers'probabilityofsuccessfullyhandling aproblem.sincewearealsointerestedinthebehaviorforclosetounity,weuseabase thesolvers.evenextendingto100,wefailtocapturethecompleteperformancedata proles.inthisway,wecanshowallactivitythattakesplacewith<mandgraspthe forlancelotandloqo.asanaloption,wedisplayalogscaleoftheperformance Wehaveseenthatatleasttwographsmaybeneededtoexaminetheperformanceof thegraphisnotasintuitive,sinceweareusingalogscale. infigure4.3.thisgraphrevealsallthefeaturesoftheprevioustwographsandthus capturestheperformanceofallthesolvers.thedisadvantageisthattheinterpretationof 7!1npsizefp2P:log2(p;s)g 1 Log 2 Scaled Performance Profile on Subset of COPS 0.9 P( log 2 ( t p,s / min {t p,s : 1 s n s } ) τ) 0.8 0.7 0.6 0.5 interval[0;log2(1043)]infigure4.3toincludethelargestp;s<m.weextendtherange 0.4 0.3 theothertwoguresand,inaddition,showsthateachofthesolversfailsonatleast8%of slightlytoshowtheatliningofallsolvers.thenewgurecontainsalltheinformationof Figures4.1and4.2aremappedintoanewscaletoreectalldata,requiringatleastthe Figure4.3:Performanceproleinalog2scale LANCELOT 0.2 MINOS 8 SNOPT LOQO 0.1 0 0 2 4 6 8 ρ M τ

problemsweregenerallychosentobedicult. theproblems.thisisnotanunreasonableperformanceforthecopstestsetbecausethese 5CaseStudy:TheFullCOPS 1 Log 2 Scaled Performance Profile on Full COPS Set 0.9 P( log 2 ( t i,j / min {t i,j : 1 j n s } ) τ) 0.8 0.7 0.6 0.5 0.4 0.3 theperformanceprolesinsection4,thisgureshowsthatperformanceproleseliminate Figure5.1:PerformanceproleforfullCOPSset 0.2 theundueinuenceofasmallnumberofproblemsonthebenchmarkingprocessandthe LANCELOT sensitivityoftheresultsassociatedwiththerankingofsolvers.inaddition,performance Wenowexpandouranalysisofthedatatoincludealltheproblemsinversion2.0ofthe MINOS 0.1 SNOPT prolesprovideanestimateoftheexpectedperformancedierencebetweensolvers. COPS[7]testset.WepresentinFigure5.1alog2scaledviewoftheperformanceproles forthesolversonthattestset. Figure5.1givesaclearindicationoftherelativeperformanceofeachsolver.Asin LOQO 0 0 2 4 6 8 10 ρ testsetloqodominatesallothersolvers:theperformanceproleforloqoliesabove ThemostsignicantaspectofFigure5.1,ascomparedwithFigure4.3,isthatonthis M τ allothersforallperformanceratios.theinterpretationoftheresultsinfigure5.1is important.inparticular,theseresultsdonotimplythatloqoisfasteroneveryproblem. Theyindicateonlythat,forany1,LOQOsolvesmoreproblemswithinafactorof ofanyothersolvertime.moreover,byexaminingps(1)andps(m),wecanalsosaythat 9

LOQOisthefastestsolveronapproximately58%oftheproblems,andthatLOQOsolves themostproblems(about87%)tooptimality. Figure5.1withFigure4.3andnotingthattheperformanceprolesofMINOSandSNOPT SNOPTtodeteriorateonthefullCOPSset.Thisdeteriorationcanbeseenbycomparing setismuchlargerthanfortherestrictedsubsetofoptimalcontrolandparameterestimation problems.since,asnotedinsection3,minosandsnoptaredesignedforproblemswith amodestnumberofdegreesoffreedom,weshouldexpecttheperformanceofminosand factors.firstofall,ascanbeseenintable3.1,thedegreesoffreedomforthefullcopstest ThedierencebetweentheresultsinSection4andtheseresultsisduetoanumberof thatminosandsnoptuseonlyrst-orderinformation,whileloqousessecond-order information.thebenetofusingsecond-orderinformationusuallyincreasesasthenumber ofvariablesincreases,sothisisanotherfactorthatbenetsloqo. aconvenienttoolforcomparingandevaluatingtheperformanceofoptimizationsolvers,but, aresimilarbutgenerallylowerinfigure5.1. likealltools,performanceprolesmustbeusedwithcare.aperformanceprolereects Theresultsinthissectionunderscoreourobservationthatperformanceprolesprovide AnotherreasonforthedierencebetweentheresultsinSection4andtheseresultsis theperformanceonlyonthedatabeingused,andthusitisimportanttounderstandthe testsetandthesolversusedinthebenchmark. casestudy,weusedataobtainedbymittelmann[13].figure6.1showsaplotoftheperformanceproleforthetimeratiosinthedatabenchmarkoflpsolversonalinux-pmanceprolesaremostusefulincomparingseveralsolvers.becauselargeamountsofdata aregeneratedinthesesituations,trendsinperformanceareoftendiculttosee.asa 6CaseStudy:LinearProgramming (5-25-2000),whichincludesresultsforCOPLLP-1.0,PCx-1.1,SOPLEX-1.1,LPABO-5.6, Performanceprolescanbeusedtocomparetheperformanceoftwosolvers,butperfor- withoutconvergenceunderthesolver'sstoppingcriteria.onefeatureweseeinthegraphof Mittelmann'sresultsthatdoesnotappearintheCOPSgraphsisthevisualdisplayofsolvers thatneveratline.inotherwords,thesolversthatclimbothegrapharethosethatsolve allofthetestproblemssuccessfully.aswithfigure4.3,alloftheeventsinthedatatinto thosesolvesthataremarkedintheoriginaltableasstoppingclosetothenalsolution MOSEK-1.0b,BPMPD-2.11,andBPMPD-2.14. thislog-scaledrepresentation.whilethisdatasetcannotbeuniversallyrepresentativeof InkeepingwithourgraphingpracticeswiththeCOPSset,wedesignateasfailures particular,thetestsetusedtogeneratefigure6.1includesonlyproblemsselectedby benchmarkingresultsbyanymeans,itdoesshowthatourreportingtechniqueisapplicable beyondourownresults. oftheperformanceoflpsolversonlyonthedatasetusedtogeneratetheseresults.in AsinthecasestudiesinSections4and5,theresultsinFigure6.1giveanindication 10

1 Log 2 Scaled Performance Profile for LP Solvers 0.9 BPMPD 2.14 P( log 2 ( t p,s / min {t p,s : 1 s n s } ) τ) 0.8 MOSEK 0.7 PCx 0.6 SOPLEXC LPABO 0.5 BPMPD 2.11 Mittelmannforhisbenchmark.Theadvantageoftheseresultsisthat,unlikethesolversin COPL_LP 0.4 Sections4and5,allsolversinFigure6.1havethesamerequirements. 0.3 7Conclusions Figure6.1:Performanceproleforlinearprogrammingsolvers 0.2 Wehaveshownthatperformanceprolescombinethebestfeaturesofothertoolsforbenchmarkingandcomparingoptimizationsolvers.Clearly,theuseofperformanceprolesisnot 0.1 restrictedtooptimizationsolversandcanbeusedtocomparesolversinotherareas. 0 ThePerlscriptperf.plontheCOPSsite[7]generatesperformanceprolesformatted 0 2 4 6 8 10 12 ρ M τ adjustedwithinmatlabtoreectparticularbenchmarkinginterests. toshowthefullareaofactivity.theareadisplayedandscaleofthegraphcanthenbe asmatlabcommandstoproduceacompositegraphasinfigures4.1and4.2.thescript acceptsdataforseveralsolversandplotstheperformanceproleonanintervalcalculated 11

Acknowledgments Version1.0oftheCOPSproblemswasdevelopedbyAlexanderBondarenko,DavidBortz, LizDolan,andMichaelMerritt.Theircontributionswereessentialbecause,inmanycases, References thewayforcopswithhisbenchmarkingwork. ouslysharedtheiramplexpertisewithus.finally,wethankhansmittelmannforpaving version2.0oftheproblemsarecloselyrelatedtotheoriginalversion. andspiriteddiscussionsonproblemformulation,whilebobfoureranddavidgaygener- AlexBondarenko,NickGould,SvenLeyerandBobVanderbeicontributedinteresting [1]H.Y.Benson,D.F.Shanno,andR.J.Vanderbei,Interior-pointmethodsfor [3]A.S.Bondarenko,D.M.Bortz,andJ.J.More,COPS:Large-scalenonlinearlyconstrainedoptimizationproblems,TechnicalMemorandumANL/MCS-TM-237, ArgonneNationalLaboratory,Argonne,Illinois,1998(RevisedOctober1999). large-scalemixedcomplementarityproblems,comp.optim.appl.,7(1997),pp.3{25. nicalreportorfe-00-02,princetonuniversity,princeton,newjersey,2000. [2]S.C.Billups,S.P.Dirkse,andM.C.Ferris,Acomparisonofalgorithmsfor nonconvexnonlinearprogramming:jammingandcomparativenumericaltesting,tech- [5]A.R.Conn,N.I.M.Gould,andP.L.Toint,LANCELOT,no.17inSpringer [6],NumericalexperimentswiththeLANCELOTpackage(ReleaseA)forlarge-scale [4]I.Bongartz,A.R.Conn,N.I.M.Gould,M.A.Saunders,andP.L.Toint, AnumericalcomparisonbetweentheLANCELOTandMINOSpackagesforlarge-scale numericaloptimization,report97/13,namuruniversity,1997. [8]E.D.DolanandJ.J.More,BenchmarkingoptimizationsoftwarewithCOPS, [7]COPS.Seehttp://www.cops.com/. SeriesinComputationalMathematics,Springer-Verlag,1992. TechnicalMemorandumANL/MCS-TM-246,ArgonneNationalLaboratory,Argonne, Illinois,2000. nonlinearoptimization,math.programming,73(1996),pp.73{110. [10]P.E.Gill,W.Murray,andM.A.Saunders,SNOPT:Analgorithmforlargescaleconstrainedoptimization,ReportNA97-2,UniversityofCalifornia,SanDiego, MathematicalProgramming,TheScienticPress,1993. [9]R.Fourer,D.M.Gay,andB.W.Kernighan,AMPL:AModelingLanguagefor 1997. 12

[11]N.J.Higham,AccuracyandStabilityofNumericalAlgorithms,SIAM,Philadelphia, [12]H.Mittelmann,BenchmarkinginteriorpointLP/QPsolvers,Opt.Meth.Software, Pennsylvania,1996. [14]B.A.MurtaghandM.A.Saunders,MINOS5.4user'sguide,ReportSOL83-20R, [13],Benchmarksforoptimizationsoftware,2000.Seehttp://plato.la.asu.edu/ bench.html. StanfordUniversity,1995. 12(1999),pp.655{670. [17]R.J.VanderbeiandD.F.Shanno,Aninterior-pointalgorithmfornonconvex [16]R.J.Vanderbei,LOQOuser'smanual{Version3.10,Opt.Meth.Software,12 [15]S.G.NashandJ.Nocedal,AnumericalstudyofthelimitedmemoryBFGSmethod (1999),pp.485{514. andthetruncatednewtonmethodforlargescaleoptimization,siamj.optim.,1 (1991),pp.358{372. nonlinearprogramming,comp.optim.appl.,13(1999),pp.231{252. 13