Optimal processor allocation for sort-last compositing under(4)

No.of

renderers

16

64

128CompositingprocessorsNumberTimeNumberTimeNumber

Time1150.24s630.24s1270.25s2100.24s420.26s850.26sMaximumnumberofcompositingoperationsperprocessor3581315214463752210.46s0.46s1.56s1.56s3.11s2721994210.46s0.47s1.35s1.35s3.80s6.91s13.8s554218189440.46s0.69s1.57s1.58s3.13s7.34s7.34s80100214s214s

Table1.Numberofcompositingprocessorsusedandcompositingtimeafter rstimage(toignorepipelinestartupoverhead)forafullBSP-treewith16,64,or128renderingprocessors.

SL-fullversusSL-sparse:NotethatonecaneasilychangetheSL-fullimplementedabovewithaSL-sparse.Allthecom-plexityofmanipulatingsparseimagescanbelocalizedinsidethefunctionsthatsendandreceivedtheimages.Formaximumperformanceand exibility,therenderingalgorithmsshouldgenerateimagesinscan-lineorder,andalsoprovideacompactrepresentationforscan-lines(i.e.,onlythefullpixelsarerepresented).

5.PERFORMANCERESULTS

Forourexperiments,weusedanIntelParagonXP/SrunningSUNMOS(installedatSandiaNationalLaboratories),andweuseditinNXcompatibilitymode.WestudiedthemaximumframeratethatcanbeachievedwithourpipelinedevaluationschemeforagivenfullBSP-tree,andhowthisframeratedegradesasweincrease(i.e.,themaximumnumberofcompositingoperationsperformedononecompositingprocessor).Werantestsoverthreedifferentrenderingcon gurations:16,64and128renderingprocessors;andseveralvariationson,leadingtoseveralcompositingcon gurations.Notethatbyincreasingfromto(thenumberofcompositingoperations)thenumberofcompositingprocessorsdecreasesfromto.Table1summarizestheperformanceoneachofthesecon gurations;Figure5showsgraphicallyhowthecompositingtimechangesasincreases.Inthesetests,ourprimaryinterestwastostudythecorrelationofandthecompositingtime.Toachievethecorrecteffect,weneedtomakethecompositingclusteroperateatitsmaximumspeed(foragiven).Thiswasdonebymakingtherenderingprocessorsrenderasingleimage,andsimplysendthesameimageoneverysubsequentrequest.Thisisthescenariowherecompositingclusteristhebottleneckoftherenderingprocess.

ThetimesreportedintheTable1arethosereportedbythePVRcollectornode,andrepresentactualwall-clocktimes.Thatis,ifrenderingwasfastenough,andtheimagescouldbepushedonaframebufferbythecollector,itwouldbetheactualframerateauserwouldget.Inparticular,nothingelseneedstobedonetotheimagestopreparethemforpresentation;infact,thefullycomposedimageisstoredonthenodethatcontaintherootofthecompositingtree.Notethatframe-to-framecoherencedoesnotmatter,sincewearenotexploitingasparseimagerepresentation(theseexperimentsonlyshowtheperformanceofaSL-fullarchitecture).

Thecompositingcapacityrequiredofacompositingpipelineisde nedasthenumberofframesthatneedtobecomposedperunittime.Notethatbyvaryingthenumberofrenderingnodesfrom16to128,essentially,wemakethecompositingtreeworkharder.With16renderingnodesproducingimagesat4framespersecond,weneedacompositingcapacityof64framespersecondinourcompositingpipeline.TheIntelParagonhasveryslowprocessorsbytoday’sstandards,actuallyusingourimagerepresentation(anRGBaimageisstoredasfour oatsperpixel),ittakes0.22secondstoalphacompositetwo250250images.So,asingleprocessorcancomposite4.5framespersecond.Toobtainacompositingpipelinewithcapacityof64framesasecondneeds14.2(64/4.5)processors.Hence,inourexperimentalsetup,thecompositingpipelineformsthebottleneckevenatthelowestrenderingspeeds(16renderingnodesat4framesasecond).

ThefollowingobservationscanbedrawnfromthedatainTable1:

Asincreases,theframeratesdecreaseaccordingly,sincethecompositingcapacitydecreases.Alsowecanseethatourtreepartitioningschemeiseffectiveindistributingtheload.Onecanseeclearlythe niteboundarieswhereitispossibletosaveaprocessorandstillachievethesameframerate(e.g.,whenisequalto8or13).Actuallywithourpartitioningalgorithm,onecanreliablypredicttheframeratebased(almost)solelyonthecompositingcapacityofagivencompositingtree.

Inourmeasurements,thiswasaccomplishedbymakingComputeImagereturnapre-computedimageimmediatelyuponcall.

In this paper, we consider a parallel rendering model that exploits the fundamental distinction between rendering and compositing operations, by assigning processors from specialized pools for each of these operations. Our motivation is to support the para

16

14

12

Time in seconds1086

4

2

001020304050K6070809010016 Rendering ProcessorsFigure5.Variationofoverallcompositingtimewithacrossthethreebenchmarkedcon gurations.Noticethatthecom-positingcapacityneededincreaseswiththenumberofrenderingnodes.Inordertokeepthedesiredframerate,oneneedstoincreasethenumberofprocessorsallocatedtothecompositingtree.Bykeepingconstant,thisisachievedautomatically,sincethenumberofcompositingprocessorsneededalsogrow,andcanbecomputedbyouroptimalpartitioningalgorithm.

Ourasynchronousevaluationofthecompositingtreehidesalmostallthecommunicationcost.Infact,theframeratesofthepipelineareindependentofitsdepth.Furthermore,theoverallspeedofthepipelineisdirectlyrelatedtothemaximumnumberofcompositionsperformedbyeachnode(relatedto).Forinstance,wheneveryprocessorperformsonecompositingoperation,theframerateis4framespersecond(i.e.,0.25sperimage),extremelyclosetothebestachievableframerateof4.5.Ascanbeseenfromthedata,italsodegradesgracefully.

6.RELATEDWORK

Mostparallelrenderingwork(forbothgeometricandvolumetricprimitives)ongeneralpurposeMIMDmachineshaveusedthesameprocessorsforbothphases.Infact,manytechniqueshavebeendevisedtoeffectivelyinterleavethetwophasesononeprocessor.Forexample,involumerenderingusingtheBinary-Swapmethod,allprocessessynchronizebetweenrenderingandcompositingphasesaswellasduringcomposition.Forpolygonrendering,themethoddescribedbyEllsworthchangesstateslocallybetweenthetransformationandtherasterizationphases,avoidingglobalsynchronization.Incontrast,usingprocessorsperformspecializedtasks,therenderingandcompositingphasescanoverlapintime,andinfact,canbepipelined.Hardwarebuildershavebeenusingdualtypecon gurationsforalongtime.Thedistinctionbetweenthetwocategorieshasbee …… 此处隐藏：6463字，全部文档内容请下载后查看。喜欢就下载吧 ……