N 87- 26 000 F

NAS-CURRENTSTATUSANDFUTUREPLANS N 87- 26 000 F. R.Bailey NASAAmesResearchCenter MoffettField,California

INTRODUCTION derived from assuming a solution in 15 min of central-processor time are compared with the capa- The Numerical Aerodynamic Simulation (NAS) bilities of several supercomputer generations Program had its inception in 1975 at the Ames (Peterson et al., 1985). As indicated, current Research Center (ARC) when a small group of supercomputers can adequately address inviscid researchers associated with the computational flows, but the computers needed to adequately fluid dynamics program set out to obtain signifi- address more complex Reynolds-averaged approxima- cantly greater computer power and the memory tions to turbulent flows about full aircraft con- capacity needed to solve three-dimensional fluid figurations will not be available until the end of flow models (Peterson et al., 1984). The state of this decade. Large-eddy simulations for complete the art in computational aerodynamics was at the aircraft must wait for future, more powerful point where problems involving complex geometries computers. could be treated only with very simple physical Sparked by this need for increased computer models and only those involving simple two- power, the ARC team spent several years performing dimensional geometry could be treated with more requirements refinement, technical studies, and complex physical models. It was clear that to advocacy that resulted in the establishment of the treat problems with both complex three-dimensional NAS Program in 1983. The ongoing objectives of geometries and complex physics required more com- the program are: (I) to provide a national compu- puter power and memory capacity than was available. This is illustrated in figure I, where the tational capability, available to NASA, Department of Defense, and other Government agencies, indus- estimated computer speed and memory requirements try, and universities, as a necessary element in

1011 --

_.\0_'_

1010

HSP-2(PROJECTED) 109

CRAY-2

lO8 R EYNOLDS-AVERAGED NAVlER-STOKES O / ILLIAC-IV / CYBER-205, CRAY-XMP / _ 10 7 CRAY-1S /_ AC AIRCRAFT CYBER-203 / W WING CRAY-1 A AIRFOIL 106 CDC-7600 /

105

104 .001 .01 .1 1 10 102 103 104 105 106 COMPUTER SPEED, mflops

Fig. I Computer speed and memory requirements for aerodynamic calculations compared with the capability of various supercomputers.

13 ensuring continuing leadership in computational dynamics and computational physics community. The fluid dynamics and related disciplines; (2) to act NPSN will be implemented in phases, with the com- as a pathfinder in advanced, large-scale computer pletion of each phase resulting in an NPSN config- system capability through systematic incorporation uration keyed to the introduction of a new- of state-of-the-art improvements in computer hard- generation, high-speed processor. The NAS Program ware and software technologies; and (3) to provide defined the first two configurations in 1983. a strong research tool for NASA's Office of Aero- These are the Initial Operating Configuration nautics and Space Technology. (IOC) and Extended Operating Configuration (EOC) defined below. An early task in implementing the NAS Program was to ensure its role as a pathfinder in Initial Operating Configuration: providing advanced supercomputing capabilities to the computational fluid dynamics community. To High-speed processor-1 (HSP-I) accomplish this a long-term strategy of instal- 250 million floating-point ling, at the earliest opportunity, a system repre- operations/sec (MFLOPS) senting each new generation of increasingly more computing rate for optimized, powerful high-speed processors was initiated. large-scale, computational aerody- Each generation may be a prototype or early namic applications production model. The strategy, illustrated in 256 million 64-bit words of central figure 2, has already started with the installa- memory tion of Ser. No. 2 of the Cray-2 as the first gen- Integrated and expandable NPSN configuration eration, high-speed processor. Introduction of Common-user interface for all subsystems the Cray-2 required an 8-mo period of test and Medium-band (56 kilobits/sec) communica- integration before production operations began. tions for remote sites The second high-speed processor will be added in Wide-band (1.544 megabits/sec) communi- 1987 followed by several months of test and in- cations to NASA centers tegration leading to an operational capability in 1988. The third generation will replace the first Extended Operating Configuration: in the 1989/90 time frame. In subsequent years, new advanced high-speed processors will replace Additional High-Speed Processor-2 (HSP-2) the older installed systems, thereby maintaining a 1000 MFLOPS computing rate for opti- steady state in which there are two high-speed mized, large-scale, computa- processors, one of which is the most advanced tional aerodynamic applications available. Upgraded subsystems to accommodate HSP-2 and graphics subsystems Coupled with the high-speed processor instal- 6.2 Megabits/sec communications to NASA lation strategy, is the ongoing design and devel- centers opment of the NAS Processing System Network (NPSN). The NPSN is a network of high-speed processors and support computers configured to provide a state-of-the-art, scientific, supercomputing environment to the computational fluid

3RD GENERATION HIGH-SPEED PROCESSOR

TEST ] OPERATE

INTEGRATE

2ND GENERATION HIGH-SPEED

,.I t PROCESSOR TEST [ OPERATE < I/////t,,- !> Q. INTEGRATE < t_ r_ 1ST GENERATION HIGH-SPEED PROCESSOR z

< (///// TEST J OPERATE uJ INTEGRATE I ¢J z

MID 1986 1988 1989/90

TIME II IB p.-

Fig. 2 NAS Program strategy for introducing state-of-the-art high-speed processors.

14 TheNASFacility, completedin January1987, NASIOCDESCRIPTION providesthephysicallocationfor theNASPro- gram.It housestheNPSN,systemsupportand developmentstaff, andcomputationalfluid dynam- TheIOCof theNPSNconsistsof thefollowing ics researchstaff. Thisnew90,O00-ft2 building functionalsubsystems:high-speedprocessor, wasdesignedwithapproximately15,000ft 2 of supportprocessing,workstation,mass-storage,and raisedcomputerfloor whichis sufficient to long-haulcommunications.Theconfiguration, accommodatefutureexpandedNPSNconfigurations. shownin figure 3, is a local-area-network- Approximately15,000ft2 of additionalraised orientedarchitectureconsistingof hardwarefrom floor is availablefor technicalsupportspace, a numberof vendors.Toreducehardwareand laboratories,andusercolocatedworkstation vendor-specificdependencies,theNPSNsoftware equipment.Officespacewasdesignedto allowthe presentsa commonuserinterfaceacrossthesystem computersystemsstaff andcomputationalfluid that providesuserswith the sameenvironmenton dynamicsresearchersto sharethebuilding. These all user-accessiblemachines,i.e., provides accommodationsprovidea strongbondingbetween commonutilities andcommandsonuser-visible thecomputerserviceprovidersandtheir subsystems.Thisuniformenvironmentenables researcherclients. usersto moveeasilybetweenprocessors,allows for easyfile accessandcommandinitiation across TheNASSystemsDivision,formerlytheNAS machineboundaries,andenhancesuser-codepor- ProjectsOffice,at ARChascompletedtheinitial tability withintheNPSN. tasksandis responsiblefor accomplishingtheNAS Programlong-termobjectivesona continuing Theoperatingsystemonall user-visible

basis. TheDivisionincludesa skilled staff of computersis basedonAT&T'sUNIXTM System V.2 computerscientistsandengineerswhoare operating system with network software modeled responsiblefor planning,designing,andimple- after the Berkeley 4.2/4.3 bsd UNIX operating mentingtheadvancedtechnologynecessaryto keep system. Jobs on all user-visible systems can be NASattunedto its pathfindingrole. TheDivision run in batch or interactive modes. Communications alsoprovidesday-to-daycomputeroperations; protocols come from the DoD Internet family of ongoingcomputersystemsupportfunctionssuchas protocols, commonly referred to as TCP/IP. hardwareandsoftwaremaintenance;andlogistics; aswellascomputerservicespersonnelto provide thetraining, consultingandtrouble-shooting necessaryto offer premiersupercomput[ngservice to a nationaluserbase.

'_71 - COMPUTER TAPES NETWORK SUBSYS.

STORAGEDISK (300 GB) CARTRIDGE TAPE __! MICOM i__z__ TERMINALSLHCS

_-_---_---_ M!Lnet L IMP ) I VlTALINKI__7

HSP-1 / I ETHERNET . l BRIIDGE I *--1_ NASnet CRAY-2 1 II . _I -_ SPS '_r'_ /i_/ I Irx /I I_ AI III I . -- I AM?.,l AMSPASHLI c_ I V_AX 11/780 .J-J'J VkAAX GWAYSI_]_- I I I ITO

C ENA'_'CAL HSDN HYPERCHANNEL COMPUTING FACILITY

-_ _ WKS _ _ ,,_ ! I ET"ERN T|

Fig. 3 IOC of the NPSN.

15 High-SpeedProcessor Mass Storage Subsystem Thefirst HSP,a Cray-2,is a four-processor systemwith268,435,456wordsof dynamicmemory-- The Mass Storage Subsystem (MSS) consists of or asmorecommonlyreferenced,with256megawordsan Amdahl 5860 mainframe computer, Amdahl 6380 (whereeachik equals102464bits/word)of disks, and IBM 3480 cartridge tape drives; this dynamicmemory.Thefourprocessorscanoperate subsystem serves the global online and archival independentlyonseparatejobsor in a multipro- storage needs of the NPSN. The MSS checks and cessingmodeona singlejob. TheNASCray-2has coordinates requests for files to be stored or achieveda peakcomputingspeedof 1720MFLOPSfor retrieved within the NPSN and maintains a direc- assembly-languagematrixmultipliesanda sus- tory of all contained files. The MSS also acts as tainedaggregate-speedin excessof 250MFLOPSfor a file server for the other NPSN subsystems, con- a jobmixof computationalphysicsFORTRANcodes. trois its own internal devices, and performs file duplication, media migration, storage allocation, Commonmemoryis accessedautomaticallyby accounting and file management. The aggregate thehardware,andcanbeaccessedrandomlyfrom Amdahl 6380 disk-storage capacity is approximately anyof thefourprocessorsaswell asfromanyof 300 gigabytes at present. thecommondatachannels.Anyjob canutilize all or partof thecommonmemory.Thismemoryis dividedinto four quadrants,eachof whichhas128 High-Speed Data Network interleaved banks. In addition to the common memory, each of the processors has a very-high- The High-Speed Data Network (HSDN) is the speed local memory for temporary storage of vector local communications network which allows the and scalar data. Externally, there are 34 DD49 exchange of data and control messages within the disks attached to the Cray-2, giving a combined NPSN. Major design emphasis was placed on its storage capacity of approximately 40 gigabytes. ability to support large file transfers between Cray-2 software includes an operating system the subsystems. The local network includes (UNICOS) based on AT&T UNIX System V.2; an auto- HYPERchannel, Ethernet, and the driver-level net- matic vectorizing FORTRAN compiler; a C-language work software to support the intra-NAS data compiler; a large set of batch and interactive communications. utilities; a large set of libraries, including a multitasking library; TCP/IP networking; and Berkeley "r" co,_nands. Long-Haul Communication Subsystem

NPSN access by the nationwide remote user Support Processing Subsystem community is provided by the Long-Haul Communica- tion Subsystem (LHCS). LHCS consists of a NAS- The support processing subsystem (SPS) con- unique remote-connection network, NASnet, and sists of an Amdahl 5860 mainframe computer and hardware/software interfaces to existing four VAX 11/780 minicomputers. This subsystem Government-sponsored networks. supports general-purpose interactive processing for local and remote users. In addition, it pro- NASnet, shown in figure 4, is a starlike vides the interface to local and remote terminals configuration of data communication links connect- and to tapes and high-speed printers. ing remote user sites to the NPSN. The physical data links between NASA Centers are provided by NASA's Program Support Communication Network. Workstation Subsystem These links are extended to non-NASA sites via AT&T tail circuits. Except for dedicated The Workstation Subsystem (WKS) consists of 224 Kbit/sec terrestrial links to Lewis and microprocessor-based workstations which produce Langley Research Centers, all data links are graphical displays of Cray-2-generated datasets switched terrestrial circuits with a capacity of and manipulate text and data as well as perform 56 Kbit/sec. At each termination point there is a

TM small-scale computations. The IRIS TM (Integrated Vitalink TransLAN communication bridge that con- Raster Imaging System), selected as the initial nects the data circuit to an Ethernet network. NAS workstation, is manufactured by Silicon Graph- The bridges permit NPSN Ethernet packets with ics Inc. and interfaces to both Ethernet _ and remote addresses to be sent to destination hosts

Network Systems Corporation's HYPERchannel TM net- on the remote-site Ethernets. Conversely, Ether- works. A key element of the IRIS is a set of net packets on a remote site's Ethernet can be special-purpose graphics microprocessors, called a routed to NPSN Ethernet hosts. In this way, "geometry engine," which transforms and displays remote users see their own local host, i.e., three-dimensional data at rates exceeding 50,000 workstation, as just another node on the NPSN floating point coordinates/sec. Other worksta- Ethernet. Since NASnet is implemented by switch tions with TCP/IP interface capabilities are also circuit technology, the remote user literally compatible with the NPSN System. dials up the connection only when required. From the software viewpoint, the selection of the DoD Internet Protocol (TCP/IP) suite for the NPSN

16 NASnet User Sites

ODDARD

H_LL / I .OC EED LOCKHEED _--"_ ...... _ "_ _ \ _..___-.--.-_ MCDONNELL DOUGLAS "_ _ _ _,__"'_-'-"_ _"_ NORTHROP _ _ _sr_, _ ROCKETDYNE v -_ fJOHNSON \ ROCKWELL NASA SITES ' PSCN NODES , .."" USER SITES

PSCN TRUNK LINES

AT&T TAIL CIRCUIT

k •

Fig. 4 NASnet configuration.

allows the most straightforward implementation of Technology (OAST) Research Centers (Ames-Moffett, NASnet since it not only supports a rich set of Ames-Dryden, Lewis and Langley) using TI applications (mail, file transfer, remote login, (1.544 mbit/sec) satellite links. etc.), but implementation also exists for a large number of various computers ranging from IBM PCs The second PSCN-provided capability is the to Cray supercomputers. In summary, NASnet, whose NASA Packet Switch System (NPSS). NPSS provides first prototype implementation was in July 1985, 9600 bit/sec terminal access to the NPSN from NASA provides the remote user with all the functional Centers. Since NPSS is also connected to GTE's capabilities of a local user of the NPSN and, in Telenet TM packet switched network, remote users outside NASA Centers can gain access to the NPSN particular, gives the ability to use all the at rates of 12OO-24OO bits/sec. interactive capabilities of powerful graphics workstations. For example, experience has shown Access by DoD research installations and many NASnet to be very effective in supporting NAS universities is provided by MILnet and ARPAnet. workstation clusters at Lewis and Langley Research Additional university access will be provided in Centers. Users at these Centers not only access the future by a newly formed network sponsored by the NPSN in the same manner as ARC users, but also the National Service Foundation (NSF) (Jennings utilize their Ethernet connected workstations to et al., 1986). This network (popularly called access computer resources at their own Centers. NSFnet) is designed to provide remote access to

In addition to NASnet, PSCN circuits support NSF-sponsored supercomputer centers by connecting two other long-haul communication capabilities. several regional network_ to a national backbone The first is the Computer Network Subsystem (CNS) circuit. One of these regional networks is the that provides batch-oriented bulk file trar_sfer NSF-funded Bay Area Regional Research Network service among NASA Office of Aeronautics and Space (BARRnet) which includes ARC, University of

17 Californiacampuses(Berkeley,Davis,San grants,contracts,andjoint programs.Theapprox- Francisco,andSantaCruz),andStanfordUniver- imateallocationof theremainingamountis 20%to sity. Whenoperationalin 1987,BARRnetwill DoD,15%to industry,5%to otherGovernmentagen- provideTI communicationslinks betweenpar- cies, and5%to universities. Onlythecostsof ticipatingmembersites andwill connectto industryproprietaryworkarereimbursedto the ARPAnetat Stanfordandto theNSFfundedSan Governmentin accordancewith a formulathat DiegoSupercomputerCenter(SDSC)througha remote includesoperatingandcapitalimprovementcosts. useraccesscomputer(RUAC)at UCBerkeley. TheNASProgramhasimplementedproject selectionandresourceallocationprocedureswhich NASCOMPUTATIONALSERVICES invite applicationsfor NASusagefromthe scientific usercommunityandprovidefor peerreview of projectsandequitableresourceallocation. TheNASProgramprovidesall userswith con- TheNASAOfficeof AeronauticsandSpaceTechnol- sulting, training,anddocumentationservices. A ogyreceivesapproximately45%of the resources, telephonehotlineis mannedaroundtheclock, to besharedamongAmes,Langley,andLewis excludingGovernmentholidays. Consultingadvice ResearchCenters.EachCentersuballocatesits is alsoavailablebycomputermailandin person shareto individualin-houseandsupported (for localandvisiting scientists). Consulting researchprojects. Theuniversityallocationis expertiseincludesNPSNoperatingsystems,compil- administeredbytheNationalScienceFoundation. ers, communications,codeconversion,codeoptimi- Overallresourceavailability is determinedbythe zation,file handling,tapeusage,workstation NASSystemsDivision. graphics,andotherareasof userinterest. Tobeselectedfor NASusage,potentialusers Trainingservicesinclude3-dayusertraining mustproposetechnicallysoundandrelevantproj- classeswhichareheldmonthlyat ARC.These ectswhichrequiretheuniquecapabilitiesoffered classescoverintroductoryUNIX,codeconversion bytheNPSN.TheNASselectioncriteria are: aids, codeoptimization,file archivalmethods, technicalquality, nationalneed,timeliness,and tapeusage,anduserresponsibilities. User suitability to NASresources.Selectioncriteria trainingclassescanbepresentedat remotesites andallocationguidelinesare periodically byspecialarrangement.NPSNusageinformationis reviewedto ensuremaximumbenefits. Theguide- alsoavailableonlinein theformof manualpages lines for usagedescribedhereapplyonlyto NAS aswell astutorials andsamplesessions. computersin anoperationalstatus. Whennew computersarebeingintegratedinto theNPSN,they In theareaof documentation,theNASUser are devotedprimarilyto testingandsoftware Guideis sentto eachaccountholder. TheUser development. Guidepresentsbasic,high-levelinformation. Otheruserdocumentsare referencedfor more Duringthe interimoperationalperiod,from in-depthinformation.TheUserGuidecontainsa July 1986throughFebruary1987,Cray-2computer completelist of Cray-2manuals,editor manuals, timewasallocatedto 123projects. Theselection referencecards,etc., whichareavailableby processis nowunderwayfor thefirst full opera- requestfor all users. tional period,whichis for I year,March1987 throughFebruary1988.Forthe interimperiod, In addition,NASSystemBulletinsare sentto Cray-2timewasdistributedovera widerangeof eachuserwhenmajorsystemschangesoccur,anda disciplinesasshownin figure 5. Notethat a monthlyNASnewsletteris sentto usersandother highpercentageof theCray-2wasallocatedto interestedpeople.ThenewsletterscontainNAS projectsrelatedto theNationalAeroSpacePlane usageinformationalongwith itemsabouttheuser (NASP).Thisreflects theimportanceof supercom- projectsandgeneralNASnewsanddevelopments. putingto this newinitiative. NASPtechnology applicationsareexpectedto increasein future years. NASUSAGE EARLYNASOPERATIONALEXPERIENCE TheNASProgramis intendedto supportpio- neeringworkin basicandappliedresearch. Approximately90%of thescientific researchper- TheNASPrograminitiated pilot operations tains to theareasof fluid dynamics,aerodynam- for a selectgroupof usersin December1985, ics, structures,materialscience,aerothermal shortlyafter theCray-2hadcompletedacceptance dynamics,andotheraeronauticsapplications. The testing. In July 1986,interimoperationsof a remaining10%includesworkin otherdisciplines completelyintegratedNPSNbegan.Useraccess suchaschemistry,atmosphericmodeling,astro- duringtheseperiodsincreasedrapidlyandappre- physics,andotherareasof interest to NASA. ciableoperationalexperiencewasgainedbyNAS About55%of theavailablecomputerresources users. supportsNASAprogramsandis usedbybothin- housestaff andorganizationsinvolvedin NASA

18 Fig. 5 Cray-2computertimeallocationdistribution basedon 123interimoperational periodallocations.

Evenwithinthecontextof pilot operations, tion or geometriccomplexityrequiredtheuseof a whichincludeda highlevel of systemdevelopment SolidStateDisk(SSD).Theinput/output(I/0) activities, initial NASusersprovidedoverwhelm- delaysassociatedwith pagingdatabetweenthe inglypositivefeedbackontheir increasedproduc- centralmemoryanddisk weresignificant and,not tivity whileusingtheNPSN.In particular,most infrequently,reached10or 20timestheactual usersindicateda significant reductionin code CPUtimefor thejob. Therefore,specialized, development/debuggingtime,start-uptimefor machine-specificcodemodificationswererequired changesin configurationgeometryor grid resolu- for efficient utilization of SSDmemory. tion, andjob turnaroundtimes,aswell as improvementsin thegraphicalanalysisof computed Forexample,a typical three-dimensional results. Reynolds-averagedNavier-Stokessimulationis constrainedto resolveless than69cubed(approx- Thereductionin codedevelopmentanddebug- imately0.3megaword)spatialpointsif thejob is gingtimecanbeattributedmainlyto theavaila- to remainin thecentralmemoryof an8-megaword bility of interactivesupercomputingandtherich machine.Full useof 256megawordswouldallow varietyof tools in theUNIXenvironment.Prior for 217cubed,or 10.22million points, to be to thedevelopmentof theUNICOSoperatingsystem, resolvedfor an in-corerunof this example vendor-suppliedoperatingsystemsfor supercompu- code. Thusthe increasedmemoryof theCray-2 ters hadremained,almostexclusively,asbatch allowsfor bothincreasedjob complexitywithout modeprocessing.Whilebatchprocessingmaybe significant codemodificationandvirtual elimina- desirablefor longproductionruns,debuggingand tion of I/0 wait time. Theneteffect is codedevelopmentarenot besthandledin this decreasedstartuptimefor increasinglycomplex mode.UNIXprovidesboththe interactivecapabil- problemsandimprovedjob turnaround. ity necessaryfor efficient codedevelopmentand debugging,andthebatchcapabilityfor production Asthespeedandmemoryof supercomputers runsvia theNetworkQueuingSystem(NQS)soft- continueto increase,largerandmorephysically ware. Forthesystemdeveloper,manyof thesame complexproblemscanbesolvedmoreroutinely. To advantagesarerealized.In addition,it hasbeen analyzetheselargerandmorecomplexdatasets, reportedthat the introductionof newsystems,as graphicaldisplaysarenotonlydesirablebut wellassoftwareenhancementsand"bugfixes,"are necessary.Specializedgraphicalprocessorssuch moreeasilyandquicklycompletedwith theuniform asthe"geometryengine"of the IRISworkstation UNIXenvironmentandconsistentprotocolsuite helpto providepseudo-real-timevisualizationof acrosstheNASmachines. computedresults, i.e., real-timemanipulationof graphicalimageswhichare createdfrompreviously TheNASCray-2is the first full-memory, computeddatasets. four-processorCray-2.The256-megawordcentral memorycapacityrepresentsa two-orders-of- It is probablyrealistic to assumethat the magnitudeincreaseoverpreviouslyavailable demandfor supercomputersandgraphicaldisplays capacity. Prior to theCray-2,increasedresolu- expandsto fill anyavailableresource.However, 19 becauseof thedistributedprocessingcapability figurations. EOC,plannedto beoperationalin of theNPSNandits uniformsoftwareenvironment 1988,addsanadvancedsupercomputer,theHSP-2, andprotocolsuite, coproeessingbetweenworksta- withperformanceandmemorycapacitytargetedat tion andsupercomputerscanbeeasily implemented fourtimesthat of theCray-2. for moreeffectiveutilization of available Theadditionof this largeincrementin com- resources. puterpowerraisesthepotentialfor inbalance Graphicalcoprocessingtasks,suchasthe amongprocessingrate, datatransportbandwidth, ARC-developedRealtimeInteractiveParticleTraces andstoragecapacityacrosstheentire NPSN.At (RIP)softwarepackagewhichinteractivelycalcu- present,theperformanceandstoragecapacityof lates particle pathsin acomputedflow field, supercomputersis outpacingtheability to move involveinitiating twocoupledprocessessimul- datato andfromthecomputerandto storeit. taneously:a computationalandmemoryintensive Newstoragetechnologiessuchasoptical, magneto- processresidingontheCray-2anda graphical optic, andmagneticvertical recordinglookprom- displayandmanipulationprocessresidingonthe ising. Unfortunately,supercomputingneedsalone workstation.Informationis passedbackandforth arenota sufficient marketforceto drivethese betweenthe twomachinesvia HYPERchanneland technologiesto thedesiredlevel of performance Ethernetlinks for remoteworkstations.In this andcapacity. Asa consequence,newstrategies way,a processingtaskcanbeintelligently with oldertechnologiesare beinginvestigatedto dividedto makeuseof thestrongpointsof the meetnear-futureneeds. variousmachinesin theNASsystem.Fortheuser, this translatesinto a substantialsavingsin time Increasedhigh-speedprocessordataproduc- for thedisplayof aparticle traceona workstation, highergraphicsdataconsumption,andhigher tion, sincetheCray-2/IRIScombinationoffers a storagecapacitywill requireincreaseddatacom- 10:1performanceimprovementoveranIRISalone. municationbandwidthin theNPSN.Ananalysisof Becauseof theuniformUNIXenvironmentacrossthe thebandwidthrequirementfor a 20-foldincrease Cray-2andworkstations,thecreationandinstal- in processingpower,basedonmodelsof theNPSN lation of RIPrequiredoneafternoonto coordinate projectedworkloadandsystemarchitecture,shows amongthreepeoplewithexistingbut decoupled aneffectivebandwidthof 100megabits/secis codes. necessary(Levinet al., 1987). Tomeetthis requirement,theNASProgramhasefforts underway Qualityfilm outputof workstation-generatedto establisha networktest bedfor hands-on displaysis providedbyModel655Dunncamerasand investigationof newnetworkingtechnologies. prototypeSeikoprinters. Userscanmakehigh qualityphotographs,transparencies,or negatives Long-hauldatacommunicationsis alsoan with theDunncameras.TheDunnsystemworks importantareafor futureimprovement.Theexpe- directly off theRUBoutputof theworkstations. riencegainedwith NASnethasdemonstratedthe For16-mmmovies,thecomputer-controlled valueof high-bandwidthandlow-delayterrestrial Model632Dunncamerafunctionswith theaid of a circuits. TheNASProgramplansto aggressively pursueopportunitiesto economicallyincrease workstationpatchpanel. bandwidthcapabilitywhenhighcapacitytranscon- Videoanimationcapabilitiesare providedby tinental fiber-opticdatatrunksbecomeavailable anABEKASdigital diskrecorder.Thissystem in thenextfewyears. In addition,anadvanced takesanRS170outputfromtheIRIS,convertsit remote-communicationsgatewayprototypeis to NTSCsignals,andrecordsit ondisk. This planned.Theobjectiveof this effort is to meter representsa significantimprovementin recording circuit capacityas throughputdemandrises and speed,editing, andspecialeffects capabilities falls. Thiswill beaccomplishedbya type-of- overoldertape-recordingdevices.Toassist the serviceroutinggatewaythat establishestheopti- userin makinganimationsequences,a program mumbandwidthanddelaycharactersto matcha calledtheGraphicsAnimationSystem(GAS)has particularapplicationservice(interactivechar- beendeveloped.GASlets theusercreatescripts acterstream,on-demanddatatransfer,bulkfile onthe IRISwhichcanbeplayedbackat a later transfer,etc.). Thebandwidthwill beoptimized timeontheanimationsystem.GASis a fairly byaddingor deletingsubchannelallocationsof "userfriendly" programin that it recordsa thebackbonecircuit andthedelayoptimizedby user'sactionsasgraphicimagesaremanipulated switchingto either low-delayterrestrial or high- ontheIRISscreen,automaticallygeneratinga delaysatellite circuits. playbackscript. If needed,theusercanedit this script to either correcttheerrorsor to add Thefirst-order softwarechallengefor the stepsto smoothouttheanimation. futureis to providethealgorithms,languages, andprogrammingtools to effectively usethe potentialof parallel architecturecomputers. FUTURECHALLENGESANDPLANS Futurehigh-speedprocessors,withouta doubt, will consistof severalif notmanyprocessors. Parallelsupercomputersareherenowin theform TheNASProgramis engagedin planningthe of thefour-processorCrayXMPandCray-2,but implementationof theEOCandits follow-oncon- their parallel multitaskingcapabilityis greatly

2O underutilized. Whenthemultitaskingof many ningandtheNASProgramwill aggressivelycon- processorsbecomesnecessaryto achievegreater tinueto developandimplementemergingsupercom- performance,muchimprovedsoftware_ill becriti- puter,communications,storage,andsoftwaretech- cally needed.TheNASProgramis placinggreater nologiesto strengthencomputationsasa critical emphasisonimprovedtechniquesfor multitasking elementin supportingtheNation'sleadershiprole andwill aggressivelypursueresearchin massively in aeronautics. parallelarchitectures. Futuretechnologywill alsosupportmore REFERENCES parallelismacrosscomputerboundariesasin the caseof coprocessingbetweensupercomputersand workstations.Special-purposeprocessors, Jennings,D.M.; Landweber,L. H.; Fuchs,I. H.; designedto performverywell ona pieceof the Farber,D. J.; andAdrion,W.R.: Computer total problem,connectedto veryfast networksare Networkingfor Scientists,Science,Vol. 231, a fertile areafor researchanddevelopment.The Feb.1986. NASProgramis activelyengagedin researchand advancedevelopmentin theseareas,with thegoal Levin,E.; Eaton,C. K.; andYoung,B.; Scaling of introducingtheminto theNPSNbytheendof of DataCommunicationsfor AdvancedSupercompu- thedecade. ter Network.Tobepres.to theIFIP/IEEE/ITC Intern.Conf.DataCommun.SystemsandTheir SUMMARY Performance,RiodeJaneiro,June22-25,1987. Peterson,V. L.; andArnold,J. 0.: Impactof SupercomputersonExperimentation:AViewfrom TheNASprogramhasmetits first majormile- a NationalLaboratory,Proc.Amer.Soc.Eng. stone--theNPSNIOC.Theprogramhasmetits goal Symp.,Atlanta,June17-19,1985. of providinga nationalsupercomputerfacility capableof greatlyenhancingtheNation'sresearch Peterson,V. L.; Ballhaus,W.F., Jr.; andBailey, anddevelopmentefforts. Furthermore,theprogram F.R.: NumericalAerodynamicSimulation(NAS), is fulfilling its pathfinderrole bydefiningand in LargeScaleScientific Computation(Seymour implementinga newparadigmfor supercomputing V. Parter,ed.), AcademicPress,NewYork, systemenvironments.TheIOCis onlythebegin- 1984.