Presentation on theme: "Gordon Bell Bay Area Research Center Microsoft Corporation"— Presentation transcript:
1 Gordon Bell Bay Area Research Center Microsoft Corporation More Wheels of Reincarnation Or A New PC+, www+ Era? Infinite processing, memory, and bandwidth @ zero costGordon BellBay Area Research CenterMicrosoft Corporation
2 The Highly Probable Future c2025 83 items from J. Coates, Futurist, Vol. 84, 1994 8.4 B, english speaking, personally tagged & identified, prosthetic assisted and/or mutant, tense people who have access & control of their medical recordsEverything will be smart, responsive to environment.Sensing of everything… challenge for science & engineering!Fast broadband networkSmart appliances & AITele-all: shop, vote, meet, work, etc.Robots do everything, but there may be conflict with labor…A “managed”, physical and man-made worldReliable weather reports“Many natural disasters e.g. floods, earthquakes, will be mitigated, controlled or prevented”Nobel prize to “economist” for “value of information”No surprises. We can see 10 years, but not 20!
5 BIO INTELLIGENCE AGE AGRICULTURAL INDUSTRIAL TECHNOLOGY DEVELOPMENT INFORMATIONCONSUMER ACCEPTANCETECHNOLOGY DEVELOPMENTBIOINTELLIGENCE2000 BC1500180019002000 ADTIME (year)R. Satava 29 July 99
6 The only thing that matters at the end of the day is, it’s a great building.
7 PC At An Inflection Point Non-PC devices and InternetPCs
8 Consumer PCsMobileCompanionsTV/AVThe Dawn Of The PC-Plus Era, Not The Post-PC Era… devices aggregate via PCs!!!Household ManagementCommunicationsAutomation & Security
9 Copyright 1999 Microsoft Corporation PCTV a.k.a. MilliBillg Using PCs to drive large screens e.g. tv sets, Plasma PanelsCopyright 1999 Microsoft Corporation
10 Another big bang? Internet to TV and audio: The Net, PC meet the TV “milliBill”SettopboxAnalog/digital cable distributionHome CATVEthernet Home networkBasic ideas:1. PC records or plays thru video cable channels.2. PC “broadcasts” art images, webcams, presentations, videos, DVDs, etc.3. Ethernet not cable?Video capturePC broadcasts are mixed into home CATV in analog and/or MPEG digital
11 Images from: http://www.nextmonet.com A gallery that sells art on line
13 The Next Convergence POTS connects to the Web a. k. a The Next Convergence POTS connects to the Web a.k.a. Phone-Web GatewaysWeb ServerPSTNTheWebVoice to WEBBridgeDataBase
14 PC will prevail for the next decade as the dominant platform… its COTS or COTS’ AND www! Moore’s Law increases performance; and alternatively reduces pricesPC server clusters with low cost OS beat proprietary switches, smPs, and DSMsHome entertainment & control …Very large disks (1TB by 2005) to “store everything” personalScreens to enhance useLack of last mile bandwidth to move pictures, data, and interact favors home mainframes aka PCsC = Commercial; C’ = Consumer
15 My betting record: No losses … so far (>5year old bets) Not TMC & MPP domination by 1995 … c1990 with Danny HillisVideo On Demand will not exist by 1995AT&T acquisition of NCR will not be successfulNot 10K by 1/2001Not 1 B internet users by 1/2001 or 1/2002Cars won’t drive themselves by 2005PCs continue with 2 digit growth through 2002
16 Outline Future predictions… 2020 and the world Caveat: How far out can we see? WWW just >5 years oldBackground: Bell-Gray c1995 our bet on SNAPMy own history of supercomputing… from last SalishanThe hardware scene in 5-10 years?Processing and Moore’s LawNetworkingDisksChallenges:OSSCommunities with dbases & hs netsASP: workbenchesIf simulation is third mode after theory, expt, what is 4th? connection with the experimental world for data; then control… biologist workbench where work is being done.
17 SNAP … c1995 Scalable Network And Platforms A View of Computing in We all missed the impact of WWW!This talk / essay portrays our view of computer-server architecture trends.(It is silent on the client cellphones, toasters, and gameboysThis is an early draft.We are sending a copy to you in hopes that you’ll read and comment on it.We would like to publish it in several forms: 2 hr video lecture, an kickoff article in a ComputerWorld issue that Gordon is editing,a monograph enlarged to be published within a year.January 1, 1995Gordon BellJim Gray
18 How Will Future Computers Be Built? Thesis: SNAP: Scalable Networks and PlatformsUpsize from desktop to world-scale computerbased on a few standard componentsBecause:Moore’s law: exponential progressStandardization & CommoditizationStratification and competitionWhen: Sooner than you think!Massive standardization gives massive useEconomic forces are enormous
19 Performance versus time for various microprocessors Moore's law has implications for speed because transistors are smaller and faster. Increases in microprocessor cache size and parallelism come from having more transistors. The result has been the quadrupling of speed every 3 years.This is fortunate since Amdahl posited that one megabyte of memory was needed for every million instructions per second the processor ran. Speed has increased at 60% per year since the late 1980s to keep up with larger memory chips.
20 Volume drives simple, cost to standard platforms Stand-aloneDesk topsPCsThis illustrates the power of volume production. If you assume that power increases linearly with the number of processing elements, then several different platforms can supply power. The most cost-effective is a gang of PCs. Microsoft's scalable server, Tiger, for video on demand demonstrates this. Using workstations is more expensive and a higher speed interconnect adds a marginal amount. The various multiprocessors are more expensive than LAN connected workstations. Multis cost about the same amount as massively parallel computers. The second most cost-effective platforms are the small multiprocessors that use the PC's microprocessor. Thus, smaller, high volume platforms beats larger more specialized multiprocessors.
21 The economics of operating systems and databases
22 The Virtuous Economic Cycle drives the PC industry… & Beowulf CompetitionVolumeStandardsDOJUtility/valueInnovation
23 The UNIX Trap: creating the myth of “open systems” “Standard” has meant different!VendorIX platforms have created the “downsizing” market that provides an apparent, cost reductionHardware platform vendors lock-in users with servers of proprietary UNIX dialects and unique chips to maintain margins for chip and UNIX developmentVendorIX R & D costs $1.4 - $2 billionImplied selling price $ billion for $1.4 billion, or a sales tax of 1 million UNIX units of $10,000Users hostage with client-server, database, and appsAn implicit or unconscious cartel has formed that maintains the industry status quo
24 xxThe UNIX Cartel and Tax: It’s not competitive and it introduces higher downstream costs10,000 companies maintain dialectsR & D costs $1.4 - $2 billionImplied selling price $ billion for $1.4 billion, or a sales tax of 1 million UNIX units of $10,000Cost could be reduced to $400 million for ONE UNIX, sales price for 1 million units would be $2, ,000NT sales price is $650; OS2 needs to sell for $1.2b/6mFurthermore:The downstream effects on database vendors is 40% R&D efficiency causing an implied database tax of 2.5x the sales price!The downstream effects on apps vendors is similar
25 SNAP Architecture---------- With this introduction about technology, computing styles, and the chaos and hype around standards and openness, we can look at the Network & Nodes architecture I posit.
26 Computing SNAP Environment circa 2000 Legacymainframes &minicomputersservers & termsPortablesLegacymainframe &minicomputerservers & terminalsWide-areaglobalATM networkMobileNetsNT, Windows& UNIXpersonserversLocal &global datacommworldATM† & LocalArea Networksfor: terminal,PC, workstation,& serversmulticomputers built from multiple simple, serversNT, Windows& UNIXpersonservers*Centralized& departmentaluni- & mP servers(UNIX & NT)† also mb/spt-to-pt EthernetCentralized& departmentalscalableuni- & mP servers*(NT & UNIX)???First, the network is uniform and ubiquitous. It links to and is compatible with mobile networks. The many kinds of networks need to become one -- distributed and point-to-point Local Area networks, private and public Wide Area Networks, proprietary terminal and cluster interconnects and protocols, and Plain Old Telephone Service, including the telephony switching fabric.If this isn't enough, then the cable networks using broadband, broadcast technology would also adopt ATM and inter-operate with the phone and data networks. However, let me not predicate the network on having to carry switched television, although in principle it could and may. Replacing the cable and broadcast televison network is a question of plant and equipment investment, and government regulation.TC=TV+PChome ...(CATV or ATMor satellite)NFS, database, compute, print, & communication servers*Platforms: X86PowerPC ... etc.Universal highspeed dataservice usingATM or ??A space, time (bandwidth), & generation scalable environment
27 Computing SNAP built entirely from PCs Legacymainframes &minicomputersservers & termsPortablesLegacymainframe &minicomputerservers & terminalsWide-areaglobalnetworkMobileNetsWide & LocalArea Networksfor: terminal,PC, workstation,& serversPersonservers(PCs)scalable computersbuilt from PCsPersonservers(PCs)Centralized& departmentaluni- & mP servers(UNIX & NT)Centralized& departmentalservers buit fromPCs???Here's a much more radical scenario, but one that seems very likely to me. There will be very little difference between servers and the person servers or what we mostly associate with clients.This will come because economy of scale is replaced by economy of volume. The largest computer is no longer cost-effective. Scalable computing technology dictates using the highest volume, most cost-effective nodes. This means we build everything, including mainframes and multiprocessor servers from PCs!TC=TV+PChome ...(CATV or ATMor satellite)A space, time (bandwidth), & generation scalable environment
29 In a decade we can/will have: more powerful personal computersprocessing x; multiprocessors-on-a-chip4x resolution (2K x 2K) displays to impact paperLarge, wall-sized and watch-sized displayslow cost, storage of one terabyte for personal useadequate networking? PCs now operate at 1 Gbpsubiquitous access = today’s fast LANsCompetitive wireless networkingOne chip, networked platforms e.g. light bulbs, cameras everywhere, & managed by PCs!Some well-defined platforms that compete with the PC for mind (time) and market share watch, pocket, body implant, homeInevitable, continued cyberization… the challenge… interfacing platforms and people.
41 Contributions of Beowulf An experiment in parallel computing systemsEstablished vision low cost high end computingDemonstrated effectiveness of PC clusters for some (not all) classes of applicationsProvided networking softwareProvided cluster management toolsConveyed findings to broad communityTutorials and the bookGB: Provided design standard to rally community!Standards beget: books, trained people, software … virtuous cycleCourtesy of Dr. Thomas Sterling, Caltech
42 High performance architectures timeline Vtubes Trans. MSI(mini) Micro RISC nMicr “IBM PC”Processor overlap, lookahead “killer micros”Cray era Cray1 X Y C TFunc Pipe Vector-----SMP >SMP mainframes---> “multis” >DSM?? Mmax. KSR SGI---->Clusters Tandm VAX IBM UNIX->MPP if n> Ncube Intel IBM->Local NOWand Global Networks n>10, Grid
43 High performance architectures timeline Vtubes Trans. MSI(mini) Micro RISC nMicr “IBM PC”Sequential programming---->(single execution stream e.g. Fortran)Processor overlap, lookahead “killer micros”Cray era Cray1 X Y C TFunc Pipe Vector-----SMP >SMP mainframes---> “multis” >DSM?? Mmax. KSR DASHSGI---><SIMD Vector--//Parallelization---THE NEW BEGINNINGParallel programs aka Cluster Computing <multicomputers <--MPP era------Clusters Tandm VAX IBM UNIX->MPP if n> Ncube Intel IBM->Local NOW Beowlfand Global Networks n>10, Grid
44 High performance architecture/program timeline Vtubes Trans. MSI(mini) Micro RISC nMicrSequential programming---->(single execution stream)<SIMD Vector--//Parallelization---Parallel programs aka Cluster Computing <multicomputers <--MPP era------ultracomputers 10X in size & price! 10x MPP“in situ” resources 100x in //sm NOW VLSCCgeographically dispersed Grid
50 Beyond Moore’s Law …>10 yrs Just FCB (faster, cheaper, better)… COTS will soon mean consumer off the shelfMoore’s Law and technology progress likely to continue for another decade for:processing & memory, storage, LANs, & WANs are really evolvingSystem-on-a chip of interesting sizes will emerge to create 0 cost systemsNo DNA, molecular, or quantum computers, or new storesAny displacement technology is unlikely … Carver Mead’s Law c1980 A technology takes 11 years to get establishedOn the other hand, we are on Internet time!
51 High Performance Computing Supers we knew are Japanese… we have to stay the course. We actually may win!PC will continue to erode capacity needScalability & COTS are in… but you have to roll your own else pay VendorIX taxesBeowulf is $14K/TB ( 6 x 4 x 40 GB)IBM 4000R 1 rack: 2x42 500Mhz processors, 84 GB, 84 disks $420K … still cheaper than the “big buys”$10-20K/node for special purpose vs $2K for a MACEMC, IBM at $1 million/TB; vs $14KWe should back radical experiments!
54 Growth of microprocessor performance 10000Cray T90MicrosSupers1000Cray 2Cray Y-MPCray C90AlphaRS6000/590Cray X-MPAlpha100RS6000/540Cray 1Si86010Performance in Mflop/sR20001803870.168818028780870.01199819801982198619881990199219941996
55 Albert Yu predictions ‘96 WhenClock (MHz) xMTransistors xMops , xDie (sq. in.) x
56 Processor Limit: DRAM Gap 60%/yr..DRAM7%/yr..110100100019801981198319841985198619871988198919901991199219931994199519961997199819992000CPU1982Processor-MemoryPerformance Gap: (grows 50% / year)Performance“Moore’s Law”Y-axis is performanceX-axis is timeLatencyCliché:Not e that x86 didn’t have cache on chip until 1989Alpha full cache miss / instructions executed: ns/1.7 ns =108 clks x 4 or 432 instructionsCaches in Pentium Pro: 64% area, 88% transistors*Taken from Patterson-Keeton Talk to SigMod
59 Exponential change of 10X per decade causes real turmoil! 100000100001000100$K 1010.10.018 MB1 MB256 KB64 KB16 KBTimeshared systemsSingle-user systems71
60 VAX Planning Model 1975: I didn’t believe it The model was very good1978 timeshared $250K VAXen cost about $8K in 1997!Costs declined > 20%users got lots more memory than I predictedSingle user systems didn’t come down as fast, unless you consider PDAsVAX ran out of address bits!72
61 System-on-a-chip alternatives FPGASea of un-committed gate arraysXylinx, AlteraCompile a systemUnique processor for every appTensillicaSystolic | arrayMany pipelined or parallel processorsDSP | VLIWSpecial purpose processorsTIPc & Mp.ASICSGen. Purpose cores. Specialized by I/O, etc.Intel, Lucent, IBMUniversal MicroMultiprocessor array, programmable I/oCradle
62 Cradle: Universal Microsystem trading Verilog & hardware for C/C++ UMS : VLSI = microprocessor : special systems Software : HardwareSingle part for all appsrun time via FPGA & ROM5 quad mPs at 3 Gflops/quad = 15 GlopsSingle shared memory space, cachesProgrammable periphery including: 1 GB/s; 2.5 Gips PCI, 100 baseT, firewire$4 per flops; 150 mW/Gflops
63 UMS Architecture Memory bandwidth scales with processing Must allow mix and match of applications. Design reuse is important thus scalability is a must. Resources must be balanced. Cradle is developing such an architecture which has multiple processors (MSPs) which are attached to private memories and can communicate with external devices through a Dram controller and programmable I/O.Explain architecture- Regular, Modular, Processing with Memory, High speed busMemory bandwidth scales with processingScalable processing, software, I/OEach app runs on its own pool of processorsEnables durable, portable intellectual property
65 Linus’s Law: Linux everywhere Software is or should be freeAll source code is “open”Everyone is a testerEverything proceeds a lot faster when everyone works on one codeAnyone can support and market the code for any priceZero cost software attracts users!All the developers write lots of code
66 ISTORE Hardware Vision System-on-a-chip enables computer, memory, without significantly increasing size of disk5-7 year target:MicroDrive:1.7” x 1.4” x 0.2” : ?1999: 340 MB, 5400 RPM, 5 MB/s, 15 ms seek2006: 9 GB, 50 MB/s ? (1.6X/yr capacity, 1.4X/yr BW)Integrated IRAM processor2x heightConnected via crossbar switchgrowing like Moore’s law16 Mbytes; ; 1.6 Gflops; 6.4 Gops10,000+ nodes in one rack! 100/board = 1 TB; 0.16 Tf
67 The Disk Farm? or a System On a Card? 14"The 500GB disc cardAn array of discsCan be used as100 discs1 striped disc50 FT discs....etcLOTS of accesses/secondof bandwidthA few disks are replaced by 10s of Gbytes of RAM and a processor to run Apps!!
69 Disk vs Tape At 10K$/TB disks are competitive with nearline tape. 40 GB20 MBps5 ms seek time3 ms rotate latency7$/GB for drive 3$/GB for ctlrs/cabinet4 TB/rack1 hour scanTape40 GB10 MBps10 sec pick timesecond seek time2$/GB for media 8$/GB for drive+library10 TB/rack1 week scanGuestimatesCern: 200 TB3480 tapes2 col = 50GBRack = 1 TB=20 drivesThe price advantage of tape is narrowing, andthe performance advantage of disk is growing
71 The virtuous cycle of bandwidth supply and demand Increased DemandIncrease Capacity (circuits & bw)StandardsCreate newserviceLower response timeTelnet & FTPWWWAudioVideoVoice!
72 Information Sciences Institute Microsoft QWest 744Mbps over 5000 km to transmit 14 GB ~ 4e15 bit meters per second 4 Peta Bmps (“peta bumps”) Single Stream tcp/ip throughputInformation Sciences InstituteMicrosoftQWestUniversity of WashingtonPacific Northwest GigapopHSCC (high speed connectivity consortium)DARPA
73 Map of Gray Bell Prize results Redmond/Seattle, WAsingle-thread single-stream tcp/ip via 7 hops desktop-to-desktop …Win 2K out of the box performance*New YorkArlington, VASan Francisco, CA5626 km10 hops
74 Ubiquitous 10 GBps SANs in 5 years 1Gbps Ethernet are reality now.Also FiberChannel ,MyriNet, GigaNet, ServerNet,, ATM,…10 Gbps x4 WDM deployed now (OC192)3 Tbps WDM working in labIn 5 years, expect 10x, wow!!1 GBps120 MBps(1Gbps)80 MBps5 MBps40 MBps20 MBps
75 The Promise of SAN/VIA:10x in 2 years http://www.ViArch.org/ Yesterday:10 MBps (100 Mbps Ethernet)~20 MBps tcp/ip saturates 2 cpusround-trip latency ~250 µsNowWires are 10x faster Myrinet, Gbps Ethernet, ServerNet,…Fast user-level communicationtcp/ip ~ 100 MBps 10% cpuround-trip latency is 15 us1.6 Gbps demoed on a WAN
76 How much does wire-time cost? $/Mbyte? Odlyzko, 1998 & Jim Gray Cost ($) TimeGbps Ethernet .2µ 10 ms100 Mbps Ethernet .3µ 100 msOC12 (650 Mbps) msDSL secPOTs secWireless sec
77 Modern scalable switches … also hide a supercomputer Scale from <1 to 120 Tbps1 Gbps ethernet switches scale to 10s of Gbps, scaling upwardSP2 scales from 1.2
78 Where are the challenges? Continued development based on clusters … Scalar processors need to compete with vectors. The U.S. has cast its lot with COTS!Explore radical alternatives.WWW is here. Now exploit it in every respect.Exploit OSS… though it may not be new!Telepresence & interactive communities!!!Grid as a prelude to:Application Service ProvidersPrototype biologist and chemist workbenchesCell laboratory, U. of WASloan sky survey
79 1st, 2nd, 3rd, or New Paradigm for science? Labscape
82 Labscape sensors Location tracking of people/samples multiple resolutionspassive and active tagsManual tasks (e.g., use of reagents, tools)Audio/video records, vision and indexingNetworked instruments (e.g., pipettes, refrigerators, etc.)
83 What am I willing to predict? Processing & data can be anywhere…Maui… in winter. BW is the limiter!Japan… if supers are so super else use PCsIn the disksApplication Service Providers: can we separate our data from ourselves and businesses (ying-yang of personal versus central services)The GRID e.g. biologist & chemist workbenches iff the IP doesn’t get in wayCollaboration ala astrophysics (high energy physics, math, earth sci. and any pure science if pure science continues!)OSS is the big bang for supercomputing??