SC’99: The 14th Mannheim Supercomputing Conference June 10, 1999 “looking 10 years ahead” Gordon Bell http://www.research.microsoft.com/users/gbell Microsoft What a difference 25 years and spending >10x makes! ESRDC c2002 40 Tflops. 5120 Proc. 640 Computers LLNL center 150 Mflops 7600 & Cray1 c1978 Talk plan We are at a new beginning… many views: installations, parallelism, machine intros(t), timeline, cost to get results, and scalabilities SCI c1985, the beginning: 1K processors (MPP) ASCI c1998, new beginning: 10K processors Why I traded places with Greg Papadapolous re. Clusters and SmPs Questions that users & architects will resolve New structures: Beowulf and NT eqivalent, Condor, Cow, Legion, Globus, Grid … Comments from LLNL Program manager Lessons Learned with “Full-System Mode” It is harder than you think – It takes longer than you think – It requires more people than you can believe – Just as in the very beginning of computing, leading edge users are building their own computers. Are we at a new beginning? “Now, this is not the end. It is not even the beginning of the end, but it is, perhaps, the end of the beginning.” 1999 Salishan HPC Conference from W. Churchill 11/10/1942 “You should not focus NSF CS Research on parallelism. I can barely write a correct sequential program.” Don Knuth 1987 (to Gbell) “Parallel processing is impossible for people to create well, much less debug?’ Ken Thompson 1987 “I’ll give a $100 to anyone who can run a program on more than 100 processors.” Alan Karp (198x?) “I’ll give a $2,500 prize for parallelism every year.” Gordon Bell (1987) Yes… we are at a new beginning! Based on clustered computing Single jobs, composed of 1000s of quasiindependent programs running in parallel on 1000s of processors. Processors (or computers) of all types are distributed and inter-connected) in every fashion from a collection using a single shared memory to globally disperse computers. TOP500 Systems by Manufacturers 500 Cray 450 SGI Number of Systems 400 350 300 IBM 250 TMC 200 HP intel Convex Sun 150 100 DEC Japan Inc. Compaq 50 other 0 6/93 11/93 6/94 11/94 6/95 11/95 6/96 11/96 6/97 11/97 6/98 11/98 6/99 Cray SGI IBM Convex HP Sun TMC intel Compaq DEC Japan Inc. other U. S. Tax Dollars At Work. How many processors does your center have? Intel/Sandia: 9000 Pentium Pro LLNL/IBM SP2: 3x(488x8) PowerPC LNL/Cray: 6144 P in 48x128 DSM clusters 1950 . Vtubes High performance architectures timeline 1960 . MSI(mini) 1980 . 1990 . 2000 Micro RISC nMicr “IBM PC” Sequential programming---->-----------------------------(single execution stream e.g. Fortran) Processor overlap, lookahead “killer micros” Cray era Trans. 1970 . 6600 7600 Cray1 X Y C T Func Pipe Vector-----SMP----------------> SMP mainframes---> “multis”-----------> DSM?? Mmax. KSR DASHSGI---> <SIMD Vector--//--------------Parallelization-------------------THE NEW BEGINNING----------------------Parallel programs aka Cluster Computing <--------------multicomputers <--MPP era-----Clusters Tandm VAX IBM UNIX-> MPP if n>1000 Ncube Intel IBM-> Local NOW Beowlf and Global Networks n>10,000 Grid Computer types -------- Connectivity-------WAN/LAN Netwrked Supers… SAN VPPuni DSM SM NEC super NEC mP Cray X…T (all mPv) Clusters GRID Legion T3E SGI DSM Mainframes Condor SP2(mP) clusters & Multis BeowulfNOW SGI DSM WSs PCs NT clusters Technical computer types WAN/LAN Netwrked Supers… New SAN DSM SM NEC mP NEC super Old Cray X…T T series World (all mPv) world: VPPuni Clustered GRID ( one Computing Legion SGI DSM program Mainframes (multiple program SP2(mP) Condor clusters & Multis NOW stream) streams) Beowulf SGI DSM WSs PCs Technical computer types WAN/LAN SAN DSM SM Netwrked Supers… MPI, NEC mP NEC super Vectorize Cray X…T Linda, PVM, VPPuni Parallellelize T series (all mPv) ??? LegionDistributed SGI DSM Mainframes SP2(mP) Condor clusters Parallellelize & Multis Computing NOW GRID Beowulf SGI DSM WSs PCs Technical computer types: Pick of: 4 nodes, 2-3 interconnects SAN DSM NEC Fujitsu Hitachi IBM ?PC? SGI cluster SGI DSM Beow/NT T3 HP? SMP NEC super Cray ??? Fujitsu Hitachi HP IBM Intel SUN plain old PCs Bell Prize and Future Peak Tflops (t) 1000 100 10 Petaflops study target 1 NEC 0.1 CM2 0.01 0.001 XMP NCube 0.0001 1985 1990 1995 2000 2005 2010 SCI c1983 (Strategic Computing Initiative) funded by DARPA in the early 80s and aimed at a Teraflops! Era of State computers and many efforts to build high speed computers… lead to HPCC Thinking Machines, Intel supers, Cray T3 series Humble beginning: “Killer” Micro? In 1981… did you predict this would be the basis of supers? SCI (c1980s): Strategic Computing Initiative funded ATT/Columbia (Non Von), BBN Labs, Bell Labs/Columbia (DADO), CMU Warp (GE & Honeywell), CMU (Production Systems), Cedar (U. of IL), Encore, ESL, GE (like connection machine), Georgia Tech, Hughes (dataflow), IBM (RP3), MIT/Harris, MIT/Motorola (Dataflow), MIT Lincoln Labs, Princeton (MMMP), Schlumberger (FAIM-1), SDC/Burroughs, SRI (Eazyflow), University of Texas, Thinking Machines (Connection Machine) Those who gave their lives in the search for parallelism Alliant, American Supercomputer, Ametek, AMT, Astronautics, BBN Supercomputer, Biin, CDC, Chen Systems, CHOPP, Cogent, Convex (now HP), Culler, Cray Computers, Cydrome, Dennelcor, Elexsi, ETA, E & S Supercomputers, Flexible, Floating Point Systems, Gould/SEL, IPM, Key, KSR, MasPar, Multiflow, Myrias, Ncube, Pixar, Prisma, SAXPY, SCS, SDSA, Supertek (now Cray), Suprenum, Stardent (Ardent+Stellar), Supercomputer Systems Inc., Synapse, Thinking Machines, Vitec, Vitesse, Wavetracer. What can we learn from this? SCI: ARPA-funded product development failed. No successes. Intel prospered. ASCI: DOE-funded product purchases creates competition First efforts in startups… all failed. – – – – Too much competition (with each other) Too little time to establish themselves Too little market. No apps to support them Too little cash Supercomputing is for the large & rich … or is it? Beowulf, shrink-wrap clusters; NOW,Condor, Legion, Grid, etc. 2010 ground rules: The component specs 2010 component characteristics 100x improvement @60% growth Chip Density Bytes/chip On chip clock Inter-system clock Disk Fiber speed (1 ch) 500. Mt 8. GB 2.5 GHz 0.5 1. TB 10. Gbps Computer ops/sec x word length / $ 1.E+09 doubles every 1.0 1.E+06 .=1.565^(t-1959.4) 1.E+03 y = 1E-248e0.2918x 1.E+00 1.E-03 1.E-06 1880 doubles every 2.3 doubles every 7.5 1900 1920 1940 1960 1980 2000 Processor Limit: DRAM Gap “Moore’s Law” 100 10 1 µProc 60%/yr. . Processor-Memory Performance Gap: (grows 50% / year) DRAM DRAM 7%/yr.. CPU 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 Performance 1000 • Alpha 21264 full cache miss / instructions executed: 180 ns/1.7 ns =108 clks x 4 or 432 instructions • Caches in Pentium Pro: 64% area, 88% transistors *Taken from Patterson-Keeton Talk to SigMod Gordon B.& Greg P.: Trading places. Or why I switched from SMPs to clusters “Miles Law: where you stand depends on where you sit.” 1993 GB: SMP and DSM inevitability after 30 years of belief in/building mPs GP: multicomputers ala CM5 2000+ GB: commodity clusters, improved log(p) GP: SMPs => DSM GB with NT, Compaq, HP cluster AOL Server Farm WHY BEOWULFS ? • best price performance • rapid response to technology trends • no single-point vendor • just-in-place configuration • scalable • leverages large software development investment • mature, robust, accessible • user empowerment • meets low expectations created by MPPs from Thomas Sterling IT'S THE COST, STUPID • $28 per sustained MFLOPS • $11 per peak MFLOPS from Thomas Sterling Why did I trade places i.e. switch to clustered computing? Economics: commodity components give a 10-100x advantage in price performance – Difficulty of making large SMPs (and DSM) – Single system image… clearly needs more work SMPs (and DSMs) are NOT scalable in: – – – – Backplane connected processors (incl. DSMs) vs board-connected processors size. All have very lumpy memory access patterns reliability. Redundancy and ft is required. cross-generation. Every 3-5 years start over. Spatial. Put your computers in multiple locations. Clusters are the only structure that scales! Technical users have alternatives (making the market size too small) PCs work fine for smaller problems “Do it yourself clusters” e.g. Beowulf works! MPI, PVM, Linda: programming models don’t exploit shared memory… are they lcd? ISVs have to use lcd to survive SMPs are expensive. Parallelization is limited. Clusters required for scalabilities or apps requiring extra-ordinary performance ...so DSM only adds to the already complex parallelization problem Non-U.S. users buy SMPvectors for capacity for legacy apps, until cluster-ready apps C1999 Clusters of computers. It’s MPP when processors/cluster >1000 Who ΣP.pap ΣP. T.fps #.K LLNL 3.9 P.pap G.fps 5.9 .66 6.1 .5 9.1 .3 ΣP.pap/C Σp/.C ΣMp./C G.fps # GB 5.3 8 2.5 ΣM.s TB 62 (IBM) LANL 3.1 64 128. 32 .6 2 - 0.5 2.0 4 .13 9.6 .5 8 5.12 8 9.6 128 64 1 16 8 (SGI) Sandia 2.7 (Intel) Beowulf Fujitsu 1.2 NEC 4.0 ESRDC 40 4.-16 128 16 76 Commercial users don’t need them Highest growth is & will be web servers delivering pages, audio, and video Apps are inherently, embarrassingly parallel Databases and TP parallelized and transparent A single SMP handles traditional apps Clusters required for reliability, scalabilities Questions for builders & users Can we count on Moore’s Law continuation? Vector vs scalar using commodity chips? Clustered computing vs traditional SMPv? Can MPP apps be written for scalable //lism? Cost: How much time and money for apps? Benefit/need: In time & cost of execution? When will DSM occur or be pervasive? Commodity, proprietary, or net interconnections? VendorIX (or Linux) vs NT? Shrink-wrap supers? When will computer science research & teach //ism? Did Web divert follow-through efforts and funding? What’s the prognosis for gov’t leadership, funding? The Physical Processor commodity aka Intel micros – Does VLIW work better as a micro than it did as a mini at Cydrome & Multiflow? vector processor… abandoned or reborn? multiple processors per chip or multi-threading FPGA chip-based special processors or other higher volume processors What Is The Processor Architecture? Clearly polarized as US vs Japan VECTORS OR VECTORS Comp. Sci. View Super Computer View MISC >> CISC RISC Language directed VCISC (vectors) RISC multiple pipes Super-scalar MTA Extra-Long Instruction Word Weather model performance 40 Tflops Earth Simulator R&D Center c2002 Mercury & Sky Computers - & $ Rugged System With 10 Modules ~ $100K; $1K /# 3 Scalable to several K processors; ~1-10 Gflop / Ft 10 9U Boards * 4 Ppc750’s 440 Specfp95 in 3 1 Ft (18.5 * 8 * 10.75”) … 256 Gflops/$3M Sky 384 Signal Processor, #20 on ‘Top 500’, $3M Mercury VME Platinum System Sky PPC Daughtercard Russian Elbrus E2K Who E2K Merced Clock GHz 1.2 0.8 Spec i/fp 135./350 45./70 Size mm2(.18u) 126. 300. Power 35. 60. PAP Gflps 10.2 Pin B/W GB/8 1.9 Cache (KB) 64./256 System ship Q4./2001 Computer (P-Mp) system Alternatives Node size: most cost-effective SMPs – – Now 1-2 on a single board, evolving to 4-8 Evolves based on n processor per chip Continued use of single bus SMP “multi” with enhancements for perf. & reliability Large, backplane bus based SMP provide a single system image for small systems, but not cost or space efficient for use as cluster component SMPs evolving to weak coherency DSMs “ Petaflops by 2010 ” DOE Accelerated Strategic Computing Initiative (ASCI) 1994 Petaflops Workshop c2007-2014. Clusters of clusters. Something for everyone SMP Clusters Active Memory 400 P 4-40K P 400K P 1 Tflops* 10-100 Gflops 1 Gflops 400 TB SRAM 400 TB DRAM 0.8 TB embed 250 Kchips 60K-100K chips 4K chips 1 ps/result 10-100 ps/result *100 x 10 Gflops threads 100,000 1 Tbyte discs => 100 Petabytes. 10 failures / day Petaflops Disks Just compute it at the source 100,000 1 Tbyte discs => 100 Petabytes 8 Gbytes of memory per chip 10 Gflops of processing per chip NT, Linux, or whatever O/S 10 Gbps network interface Result: 1.0 petaflops at the disks HT-MT HT-MT… Mechanical: cooling and signals Chips: design tools, fabrication Chips: memory, PIM Architecture: mta on steroids Storage material Global clusters… a goal, challenge, possibility? “ “ Our vision ... is a system of millions of hosts… in a loose confederation. ” Users will have the illusion of a very powerful desktop computer through which they can manipulate objects. ” Grimshaw, Wulf, et al “Legion” CACM Jan. 1997 Utilize in situ workstations! NoW (Berkeley) set sort record, decrypting Grid, Globus, Condor and other projects Need “standard” interface and programming model for clusters using “commodity” platforms & fast switches Giga- and tera-bit links and switches allow geo-distributed systems Each PC in a computational environment should have an additional 1GB/9GB! In 2010 every organization will have its own petaflops supercomputer! 10,000 nodes in 1999 or 10x over 1987 Assume 100K nodes in 2010 10 Gflops/10GBy/1,000 GB nodes for low end c2010 PCs Communication is first problem… use the network that will be >10 Gbps Programming is still the major barrier Will any problems or apps fit it? Will any apps exploit it? The Grid: Blueprint for a New Computing Infrastructure Ian Foster, Carl Kesselman (Eds), Morgan Kaufmann, 1999 Published July 1998; ISBN 1-55860-475-8 22 chapters by expert authors including: – – – – – – – – – – – – – – Andrew Chien, Jack Dongarra, Tom DeFanti, Andrew Grimshaw, Roch Guerin, Ken Kennedy, “A source book for the history Paul Messina, of the future” -- Vint Cerf Cliff Neuman, Jon Postel, Larry Smarr, Rick Stevens, Charlie Catlett John Toole and many others http://www.mkp.com/grids The Grid “Dependable, consistent, pervasive access to [high-end] resources” Dependable: Can provide performance and functionality guarantees Consistent: Uniform interfaces to a wide variety of resources Pervasive: Ability to “plug in” from anywhere 2004 Computer Food Chain ??? Mainframe Vector Super Massively Mini Parallel Processors Portable Computers Networks of Workstation/PCs Dave Patterson, UC/Berkeley Summary 1000x increase in PAP has not been accompanied with RAP, insight, infrastructure, and use. We are finally at the beginning. “The PC World Challenge” is to provide “shrink-wrap”, clustered parallelism to commercial and technical communities Only becomes true if system suppliers e.g. Microsoft deliver commodity, control software ISVs must believe that clusters are the future Computer science has to get with the program Grid etc. using world-wide resources, including in situ PCs is the new idea 2010 architecture evolution High end computing will continue. Advantage SMPvector clusters – Unclear that U.S. will produce one versus “stay the course” using 10x “killer micros” Shrink-wrap clusters become pervasive. – – – SmP (m>16) will be the cluster component, including SmP-on-a chip and board “multis”. Cost-effective systems come from best nodes. Backplanes are not cost-effective I/Cs Interconnection nets, log(p), are the challenge. Apps determine whether clusters become a general purpose versus niche structure Technical computer types: Pick of: 4 nodes, 2-3 interconnets SAN DSM NEC Fujitsu Hitachi IBM ?PC? ?PC? SGI cluster SGI DSM Beowulf ?HP? SMP NEC super Cray ??? Fujitsu Hitachi HP IBM ?PC? SUN plain old PCs The end