96
www.compaq.com Compaq Confidential - Need to Know Required Compaq AlphaServer Directions Rich Colarusso Rich Colarusso Alpha Corporate Consultant Alpha Corporate Consultant High Performance Servers Business High Performance Servers Business Unit Unit Business Critical Servers Division Business Critical Servers Division

Www.compaq.com Compaq Confidential - Need to Know Required Compaq AlphaServer Directions Rich Colarusso Alpha Corporate Consultant High Performance Servers

Embed Size (px)

Citation preview

  • Slide 1
  • www.compaq.com Compaq Confidential - Need to Know Required Compaq AlphaServer Directions Rich Colarusso Alpha Corporate Consultant High Performance Servers Business Unit Business Critical Servers Division Compaq Computer Corporation
  • Slide 2
  • Agenda Strategy & Business Update Alpha Positioning & Commitment Alpha Technology Microprocessor Roadmap System Overview Performance and Positioning Competitive Comparisons Future Directions
  • Slide 3
  • TANDEM
  • Slide 4
  • Alpha is... for Technical & Commercial Performance Numerically Intense Technical Applications #1Alpha # 2PA-RISC # 3POWER # 4,5,6MIPS, SPARC, IA-32 Commercial, General Purpose, Server Tasks #1Alpha # 2PA-RISC # 3IA-32 (Pentium) # 4,5,6 MIPS, POWER, SPARC $ $ x x x x x x x x Source: Illuminata, Inc. March 1999 If high performance is your need, Alpha has it in spades. Illuminata, 4/7/99
  • Slide 5
  • 31 November 2000 list of the 500 most powerful systems in the world High Performance Computing - TOP500 www.netlib.org/benchmark/top500.html IBM SP230 Hitachi 10 Fujitsu 9 NEC 6 SGI 6 Sun Sun 4 HP 1 Intel 1 # Systems in Top 100 Alpha
  • Slide 6
  • Market Share Gain/Loss HP 23 % + 1 IBM 16 % - 2.5 SGI 16 % - 3.4 SUN 12 % 0 Compaq 22 % + 3 Alpha is #2 Alpha is #2 and closing fast Technical Systems and Servers 1999 year in Review IDC report, Jan. 2000 Volkswagen AG & Porsche, Nissan, Ford, Ferrari mechanical design crash simulations PERFORMANCE ASCI Q U.S. DOE, Los Alamos Labs 375 GS320s - 12,000 Alpha cpus size of 5 basketball courts! 30+ TeraOPS; 100 TeraOPS by 2004! High Performance Computing #1 Sandia National Lab Cplant Cluster: 2000+ Alpha workstations largest production Linux cluster in the world Pittsburgh Supercomputing Center largest non-military computer 2700 CPUs, 6 TFLOPS start-up ahead of schedule; performance exceeding expectations APAC APAC (Australian Partnership for Advanced Computing) Largest supercomputer in the southern hemisphere Sun failed; re-bid selected Alpha Whitehead Institute, Cambridge MA Sanger Centre, Cambridge England mapping the human genome New: Geneva Proteomics, AxCell Biosciences IDC sees strong demand for technical systems; 13.5% growth/yr ATOMIC ENERGY RESEARCH CEA France JAERI - Japan MIT Laboratory of Nuclear Sciences 24%
  • Slide 7
  • Alpha Performance Wins Oscars... 1998 Best Visual Effects Titanic 1999 Best Visual Effects: What Dreams May Come 1999 Best Animated Short Film: Bunny Brian DePalmas new movie, Mission to Mars Spielbergs new movie, A. I. Animation House Blue Sky Studios Rhythm & Hues Santa Barbara Studios Industrial Light and Magic
  • Slide 8
  • worlds largest on-line bookstore NETSCAPE AlphaServers carry the load with plenty of room to spare. We had to work around all kinds of problems in other operating systems, like Suns Solaris and IBMs AIX. With Tru64 UNIX, there are no surprises. Its very reliable. We havent hit the upper end yet. That was the first sign that we had found an outstanding product. Fred Nixon, VP Engineering, MindSprind Enterprises, Inc. if youre ever afraid you might end up with a stunning annual growth rate, then you better invest in a highly scaleable computer platform... AlphaServer series knows no rival. Jeff Bezos, CEO, Amazon.com Explosive Growth the Internet for ISPs, Search Engines, Web Sites ... gives us the top-level performance required for our incredible traffic rate... performing in the thousand-hits-per-second range, an order of magnitude greater than... a year ago. The Web Team, Netscape Communication Corporation The Vatican customers directly schedule repair services over the Internet TruCluster of (3) 8400s Its a very powerful, elegant operating system. Roman Godzich, President 1999 Super Bowl Web Site ave response 4.63 sec longest running cluster: 5 years 50 million hits/day 70 of 100 largest are Alpha/Tru64 UNIX Q4: $22 million order for Alpha/Tru64 UNIX 1.2 million simultaneous users 24 8400s, 8 GB and 10 cpus each switched from SUN 3 years ago SCALABILIT Y
  • Slide 9
  • Business Strategy and Goals 1. remain the #1 PC vendor in the industry - radically simplify the PC for consumers and business - drive new forms of devices and wireless mobility 2. increase leadership in standards based servers 3. offer high-performance, inter-networked enterprise solutions - provide scaleable, high availability technology - lead in networked storage - deliver eBusiness solutions and services expertise iPAQ Aero Connectivity Suite Presario EZ2000 ProLiant 8-way 32-way IA64 TaskSmart Appliance Server Zero Latency Enterprise Alpha EV67 Microprocessor StorageWorks ESA12000 Compaq Services Network 3 millionth server recently shipped to Planet Online (UK Internet business) more than HP, IBM and Dell combined! Edge of the Internet Inspiration Technology with great technology, people can do great things Presario 1400
  • Slide 10
  • Enterprise Server Organization R&D: $1.3 B 37.2K people 32% 1999: $ 6 B CommlPCs 16% 1999: $12 B ConsumerProducts STORAGE BUSINESS UNIT Mark Lewis BUSINESS CRITICAL SERVERS Howard Elias CustomSystems Tru64 UNIX and Linux OpenVMS Alpha Technology AlphaServers NonStop Himalaya OpenVMS Tru6 UNIX 4 4 TM TANDEM NSK Q4 growth: 24% Q4 growth: 17% (2/3 from Alpha) Q4 growth: 17% The enterprise business represented 52% of Compaqs revenue in 1999. news.cpq, Jan. 26, 2000 Y2000: 50% revenue 60% profit Q4/00: 20% growth 50% revenue 90% profit Five Global Business Units Mike Winkler, VP INDUSTRY STANDARD SERVERS Mary McDowell
  • Slide 11
  • (includes services) OpenVMS ~ $ 4 B NSK ~ $ 2 B Tru64 Unix ~ $ 3 B WNT ~ $ 7 B BCSD ~ $ 9 B ISSD ~ $ 7 B volumevalue TANDEM NSK Compaq Server Revenue - 1999
  • Slide 12
  • the worlds most reliable systems NonStop Himalaya TANDEM Integrity S-Series the worlds most powerful chip Alpha EV7, 2003 Compaq Commitment to Alpha Alpha Alpha will be used by Tandem and these are machines with a life of 10 years or more. Enrico Pesatori, Compaq VP of Marketing Thats not a decision Compaq would have made if they thought Alpha was just going to be a stopgap to Merced, Jesse Lipcon, Compaq VP of High-Performance Server Division Chose Alpha vs. Merced because it is stable and mature Digital has more experience in writing translators & compilers Himalaya with Alpha will have all the same features Upgrade from MIPS to Alpha with NO CHANGES to applications Field installable upgrade to Alpha EV7 Clusters of MIPS and Alpha with ServerNet NonStop UX NonStop Kernel WHY? Pauline Nist, Tandem, VP Products & Technology
  • Slide 13
  • New Opportunities for Alpha June 7, 2000: API and Quadrics new partnership to develop high-performance Alpha supercomputers for Linux Concord, Mass., May 23, 2000 API (Alpha Processor, Inc.), a leading developer of high-performance solutions for high bandwidth and compute-intensive applications, today launched a new strategic business unit, API Networks, to leverage its expertise in the Internet infrastructure market and take advantage of the growing need for high-bandwidth hardware and software solutions. API has introduced Alpha Slot B technology which lowers production costs and requirements, effectively offering industry-standard packaging for the Alpha chip. Computer Reseller News, June 18, 1999 API & AMD: joint work on chipset to support both Alpha and Athlon and Lightning Data Transfer I/O interface Alpha Processor Inc. deomonstrated a 1 GHz Alpha chip that did not require special cryogenic cooling in June, 1999 ! This industry-first showcase of any microprocessor running 1 GHz frequency at room temperature validates our belief the boundaries of microprocessor Technology will continue to expand, especially for Alpha. Jeff Borkowski, VP of Marketing, API, February, 2000
  • Slide 14
  • In 1999, API revenue was $130 Million, selling Alpha to companies that use them in specialized, high-performance products Alpha System Vendors Who else, besides Compaq, uses Alpha?
  • Slide 15
  • SIX quarters of double-digit increases! AlphaServer Unit Sales by Quarter uni, dual and quad systems 3Q98 4Q98 1Q99 2Q99 3Q99 4Q99 100% growth! Gaining Marketshare... RISC Server Revenue Growth 1999 vs. 1998: Alpha revenue grew 8% Dataquest, July, 2000 report Total Market 20 % 10 % 37 %Alpha HP HP IBM IBM SUN SUN 30 % 40 % 31 % 23 % 21 % Q1/Q2 2000 vs. 1999 12 % Y2000 vs Y1999: Alpha revenue up 19% Q4 vs. Q3: 15% increase in GS system shipments 38% increase in orders Y2000 GS products/services: $800 Million revenue growth vs. same period prior year Q3: 30% Q4: 23% Q1/Q2: 37% Q3: 48% Q4: 30%
  • Slide 16
  • Compaq Commitment to Alpha Would we risk the worlds largest stock exchanges and telecommunication systems on an architecture were going to abandon? Tandem Systems will use Alpha New Opportunities for Alpha Continued Investment in Alpha 1999 Investments Alpha Chips: $100 million AlphaServers: $100 million 1998 200220042006... EV6EV7EV8EV9 & EV10 significant multi-year investment in new GS series Microelectronics Long Term Contractual Agreements fixed-function servers, network appliances, Internet network infrastructure Alpha has 80% Of RISC/Linux server market! IDC Q. Tracker, Q3 2000 Compaq, Samsung (API), AMD: $500 million joint investment; $1.5 billion over 5 years options to upgrade to EV7 and EV8 OpenVMS to be certified per Defense Information Infrastructure Common Operating Environment (DII COE) specifications requires explicit up to 20-year commitment U.S. DOE, Los Alamos Labs mulit-year R&D project with Los Alamos Labs and Celera Ericssons use of Alpha microprocessors will assure... a lengthened Alpha microprocessor road map... and continued access to Alpha technology for at least the next 7-10 years. IDC Bulletin #23330, November 2000
  • Slide 17
  • Integrated e-Business Solutions to NonStop Business Critical Servers High Performance, High Availability Systems TANDEM Compaqs cross-platform solution met all our requirements... with NonStop Himalaya servers continuous and reliable processing power, Alpha Servers intensive computing capabilities, and the ProLiant systems outstanding price performance. OptiMark Technologies, Securities Trading System William F Adiletta, CTO from Palmtop Enterprise Servers Internet Access Compaq is well-positioned to be considered a one-stop shop for large globally dispersed network infrastructures. Giga InformationGroup Industry Standard Systems NonStop Features ServerNet Lock Step Instructions Cluster Technology Portables, PCs, Desktops, Workstations Consumer Appliances Industry Standards PCI options, Ultra3 LVD drives, Insight Manager
  • Slide 18
  • The Most Powerful, Largest, and Longest Running Systems PERFORMANCE Alpha and Tru64 UNIX Customers longest running cluster: 5 years 28 AlphaServers largest single site: 4 billion records 250 TB of data 4 acre Data Center Demographics for 96% of US population 97 percent of all human genes are now represented in the Celera database most computing power: only the Pentagon has more! 50 million hits/day SCALABILIT Y AVAILABILITY
  • Slide 19
  • The Alpha Architecture PERFORMANCE SCALABILIT Y AVAILABILITY TOTAL COST OF OWNERSHIP L L F F E E I I X X T T L L B B Y Y I I I I
  • Slide 20
  • Alpha is a 25 Year Architecture 64 Bit CPU NW IO MEM 1000 times over 25 years Flexible incorporate new technologies Scalable in all dimensions Durable over decades
  • Slide 21
  • EV4 100 MHz 2 specint 1 cpu EV5 300 MHz 7.4 specint 1 cpu EV5 625 MHz 18.8 specint 14 cpu EV6-7 700+ MHz 40 specint 32 cpu P E R F O R M A N C E Alpha From 1992 to 2001 X 1000 X 19922001 EV6 525 MHz 30 specint 14 cpu EV6-8 1+ GHz 60 specint 32 cpu 199719951999 2000
  • Slide 22
  • By avoiding unnecessary complexity, Digital has not fallen prey to the major schedule problems encountered by other next-generation processors. Linley Gwennap, Microprocessor Report Alpha Architectural Leadership Part of Digitals success in maintaining performance leadership for three years running has been its ability to consistently meet or exceed performance targets. Linley Gwenna, Microprocessor Report Top semiconductor design engineers Innovative, patented features Sophisticated 64-bit CAD and simulation tools Fab-less Model Microelectronics $200 M in copper & SOI
  • Slide 23
  • new chip designs Multiple Implementations of the Same Architecture 1.0 2.5 5.0 EV4 Alpha Microprocessor Roadmap relative speed increase EV5 EV5-6 21164 - 250, 266, 300, 350, 375 400, 433, 440, 466, 500 190, 200, 225 EV4-5 233, 266, 275 EV6 667, 700 EV6-7 EV6-8 new process technology: CMOS.35.25.18, copper clock speeds 1000 21264 - 500, 525 600, 625 - 100, 125, 150, 166 21064 - 100, 125, 150, 166 1992 1993 1994 1995 1996 1997 1998 1999 2000
  • Slide 24
  • 166 2.95 266 5.18 500 15.0 300 7.43 350 10.1 440 13.6 600/625 30.3 525 EV6 19921995 18.8 0 5 10 15 20 25 30 35 40 40.1 667 EV67 19992000 Source: Standard Performance Evaluation Corporation SPEC - November, 2000 SPECint95 EV4 EV5 When we look at the performance category, Alpha is the undefeated heavyweight champion. Ed Schaider, The Standish Group International, Inc. EV6 Alpha has nearly continuously excelled all comers since 1992, an eternity in this industry. Illuminata, Inc. March 1999 50 45 2001 50 EV68 833
  • Slide 25
  • 100 200 300 400 500 467 316 353 442 SPARC-II 480 MHz P3-II 450 MHz R120000 400 MHz PIII 1000 MHz 444 EV67 667 MHz 379 HP 8600 552 MHz Source: Standard Performance Evaluation Corporation - SPEC - January, 2001 CINT2000 (peak) SPEC designed CPU2000 to provide a comparative measure of compute intensive performance across the widest practical range of hardware. The implementation resulted in source code benchmarks developed from real user applications. These benchmarks measure the performance of the processor, memory and compiler on the tested system. 600 700 EV68 833 MHz 663 est EV68 1000 MHz As a 64-bit chip, Alpha can process greater amounts of data per clock cycle than... 32-bit designs. This makes Alpha a faster performer, even if it has not quite reached that 1GHz barrier yet. ZDNet News, May 29, 2000 234 SPARC-III 900 MHz still not shipping mid 2001 544
  • Slide 26
  • 100 200 300 400 500 291 UltraSPARC-II 480 MHz 577 EV67 667 MHz 382 Power3-II 375 MHz 407 R120000 400 MHz 335 PIII 1000 MHz 600 369 8600 552 MHz Source: Standard Performance Evaluation Corporation - SPEC - January, 2001 CFP2000 (peak) SPEC designed CPU2000 to provide a comparative measure of compute intensive performance across the widest practical range of hardware. The implementation resulted in source code benchmarks developed from real user applications. These benchmarks measure the performance of the processor, memory and compiler on the tested system. encryption/decryption of credit card transactions encryption/decryption of credit card transactions not just for scientists... 700 900 800 EV68 833 MHz 832 est EV68 1000 MHz 482 SPARC-III 900 MHz still not shipping 658
  • Slide 27
  • 8000 160, 180 604e 233 P6 200 USII 250 R10000 200 P6 150 P6 180 Pentium II 333, 350 8200 200 8200 240 Xeon 400 R10000 250 300 USII 300 USII 336 10 20 30 40 50 450 USII 400 RS64-II 340 500 P III 450 604e 332 R12000 300 550 USII 450 8500 367 R12000 400 8500 440 600 60 1999 2000 1995 P II 266 PPCII 400 PWR3-II 375 SUN Alpha IBM HP SGI Intel SPECint95 The Alpha at the moment is still the best microprocessor on the planet. Peter ffoulkes, Dataquest Feb. 1, 1999 in CNet News Nobody is going to catch Alpha any time soon. Andrew Allison, Inside the New Computer Industry, March 1997 Since entering the RISC race in 1992, Alpha has continuously posted the best performance results. And not just by a little bit: it has dominated the high-end microprocessor performance charts. Illuminata Group EV6 EV5 300 EV5/350 EV56/400, 440, 466 EV56 625 EV56 600 EV56/500 EV67 July 99 EV68 USIII 750 PWR3-II 450 8600 550 750 1 GHZ IA-64 iTanium (Merced) USIII 900
  • Slide 28
  • 60 SPECint95 2000 1996 40 iTanium (Merced) 20011999 EV6 EV67 EV7 (new design) scalable performance with no software tuning < 50% Merceds size and < 20% Merceds cost in same process technology (.18) PLUS 1.5X THE PERFORMANCE! 2002 Microelectronics copper 75 85 EV68 (.18) EV68 (.18) EV69 (SOI) (.125) PIII 1 GHz 750 MHz 900 MHz 8600 550 The Register editor Mike Magee, 3/31/99 London Intel is now using Alphas for its IA-64 development work Source: The Associated Press Thursday, July 20, 2000 Intel Delays Shipment of New Chip EXTRA !!! Intel Corp. confirmed that all existing Pentium-based applications and operating systems will run on Merced systems, but performance for that software will be approximately the same as on a Pentium III. ComputerWorld, June 9, 1999 (applications must be tuned to use the new parallel and multitasking features)
  • Slide 29
  • EV7 Processor Interface EV7 Glueless multiprocessor via directory coherence High-speed interconnect (800 MHz RAMBUS) N E S W 4-5x RAMBUS IO
  • Slide 30
  • Single EV7 Processor System EV7 I/O RAMBUS DRAM RAMBUS DRAM EV7 + I/O + RAMBUS DRAM = SYSTEM Memory Size = up to 4GB (128 Mbit technology) Memory Size = up to 4GB (128 Mbit technology) Memory Bandwidth = 10 GB/sec Memory Bandwidth = 10 GB/sec Memory Latency = 70 ns (load to use, minimum) Memory Latency = 70 ns (load to use, minimum) PCI- X AGP I/O Port
  • Slide 31
  • EV7 the System Is the Silicon EV7 + I/O + RAMBUS DRAM = SYSTEM EV7 I/O RAMBUS DRAM RAMBUS DRAM Single EV7 Processor System EV7 I/O Large Scale EV7 System
  • Slide 32
  • 60 75 2000 1996 40 20012004 150 SPECint95 2002 EV67 EV68 (.18) EV68 (.18) EV69 (.13) 85 may not offer any clear performance advantages, Russell [G.M. HP Enterprise Systems Group] may not surpass RISC in pure computational power. says Jim Carlson, HPs marketing manager for IA-64 InformationWeek 5/18/1998 ... Intel will not be that fast within the next years and the performance crown will still be held by [Alpha]. Patrick Schmid, toms hardware guide it will take Intel at least three years to catch up with RISC vendors in price/performance Dean McCarron, Mercury Research Inc., in ComputerWorld, 6/9/99 iTanium (Merced) (.18) 800 MHz SPECint95 HPs testing: Merced 90% integer and 85% floating point of PA 8500/440 Ken Kroeker, HP Sr. Technical Consultant, Electronic Buyers News, August 25, 1999 McKinley (Merced II) McKinley 1999 EV8 EV7 PA-8900 1.2 GHz PA-8700 (.18) 720 MHz PA-8800 900 MHz "PA-RISC is a holdover from an old architecture, and it's getting minimal life support. I don't believe [HP has] any new micro-architectures planned. The company will do some process shrinks and add some RAM, and that will keep it afloat for a period of time." Kevin Krewell, MicroDesign Resources in TechWeb, Feb. 9, 2001 HP has switched foundry partners: from Intel to IBM Microelectronics
  • Slide 33
  • EV8 Processor with SMT Thread 2 Thread 3 Issue slots Thread 1 Unused time (cycles) Thread 4 looks like 4-way SMP to software; any SMP code will automatically work on SMT; multi-threaded applications will run 2-4X faster on SMT Issue slots Traditional Processor time (cycles) EV8 - Optimized for multi-stream workloads Simultaneous Multithreading provides concurrent execution of 4 threads (over 2x benefit for SORT and OLTP workloads with only 10% more real estate) multiple instruction issue pipelining single thread
  • Slide 34
  • EV8 Goals new superscalar core.13 microns, SOI CMOS 250 million transistors 512 registers 1800 pins, 150 watts 8 instructions/cycle 1400+ MHz 10 GB/sec memory BW 64 GB/sec. BW 256 in-flight instructions double performance, same GHz EV7 Goals common core with EV6.18 microns; copper 100+ million transistors 1439 pins, 100 watts, 1.5 volts up to 1200 MHz 10 GB/sec memory bandwidth 16 GB/sec. cache bandwidth Integrated memory controller (RAMBUS) Huge on-chip cache (1.75 MB L2) Massive memory bandwidth order of magnitude; 2X comml apps On-chip switches lower latencies, higher bandwidths, more cpus Lockstep mode Symmetric multi-thread execution (4 threads) double OLTP throughput EV6, 67, 68.35,.25,.18 microns 15 million transistors 152 registers 587 pins, 95-60 watts quad issue, 6-way execution 500 - 1100 MHz 2 GB/sec memory BW 4 - 5.5 GB/sec. cache BW I&D cache: 64 KB each 8 GB cache bandwidth 80 simultaneous in-flight instructions TANDEM Usparc III: 36 Xeon: 40 HP PA-8000: 56 L2 on chip cache 2+ MB Xeon: 3.2 Coupling Chip Design to System Design Alpha chip design and AlphaServer systems part of same division
  • Slide 35
  • Flagship Chip Min. Feature Clock Rate - MHz CMOS 5 CMOS 3 198919911995 Mariah NVAX/EV4 EV5 1.0 um.75 um.5 um 63 200300 0.8M1.5M9.3M 2004 EV8.13 um 1500+ 250M Wafer Size 400-600 Transistors CMOS 6 1998 EV6.35 um 500-600 FAB-6 8 15M 1999 EV67.25 um 600-800 15M 1996 EV56.35 um 9.3M Foundry FAB 2002 EV7.18 um 1100-1300 100+M 2000 EV68.18 um 1000+ 15M Microelectronics Alpha Semiconductor Technology Roadmap 1999 investment: $100 million Resources: 500+ design engineers, with 10% growth planned CMOS 4 6 FAB-4 & FAB-5 year chip ships in systems; chip availability one year earlier fully funded projects Flagship Chip Min. Feature Clock Rate - GHz 20062008 2010 EV9 EV10 EV11 0.1 um.07 um.05 um 2-3 3-44-5 0.5B1.5B5B 2025 Wafer Size 5-6 Transistors 2015 2018 2012 EV12.03 um 15B 20222020 Foundry FAB Its MORE than MHz... optimized for memory bound applications (EV7) optimized for threaded and multi-stream workloads (EV8) optimized for the structure of future systems (SMP) EV9 and EV10 already in active R&D
  • Slide 36
  • AlphaServer Family Departmental Global DS10, DS10L, ES40 Enterprise NEW GS FAMILY 1 - 32 Processors - SMP 400% more performance over todays AlphaServers Dynamic partitioning & work load balancing Hot swap & add, redundant power and console Starts $3300 $22,500$27,400 $60K-$2M 14-way 8-way 4-way DS20E dual uni GS160 GS80 GS320 32-way 16-way 8-way GS140 GS60E ~ $240K ~$550K-$2M > $80K $10,300$12,800
  • Slide 37
  • Global Solutions Enterprise Solutions Department Solutions 1-4 CPU EV6/500 EV67 667 EV67 700 EV68 1 GHz 8 MB cache Mar spring 99 ES40 GS80 GS320 Nov GS160 June Q4 EV6 466 EV67 667 1-2 CPU EV6/500 1 CPU EV6/466 early 99 DS20 mid 99 DS20E (Oct) Apr EV67 730 MHz Q3 EV6 525 GS60/140 EV67 600 EV68 833 Q1 Jan 200119992000 EV67 600 Q4 - Oct EV68 833 EV68 1 GHz EV68 1 GHz DS10L 1 - 8 CPU 1-14 CPU 1 - 32 CPU Q2 April EV68 833 DS10 EV68 833 Q2 May Q3/4 July 2002 EV68 1.2 GHz 8 MB cache Q1 AlphaServer Roadmap
  • Slide 38
  • Alpha High-Bandwidth Memory CPU MEM System elements share a bus Bus bandwidth is shared by all processors and I/O Data must be transferred in relatively small blocks Latency can vary significantly dependent upon number of outstanding transactions NEW Cross-bar topology System elements dynamically connected over multiple paths Dedicated bandwidth for large blocks of data between each pair Simultaneous connection to all CPUs Well-defined latency, limited request queueing CPU MEM Previous Bus-based topology Not Just the Fastest Processor... Switch
  • Slide 39
  • 777.9 0 200 400 600 800 1000 1200 1400 IBM 397 SUN 6001 SUN 10000 400 MHz 882.4 GS140 6/575 2.1 GB/sec bus 638.7 HP N4000 836.1 HP V2600 376.4 SGI O2000 300 388 296.1 261 McCalpin Streams Memory Bandwidth in MB/Sec Streams TRIAD 1 CPU configurations specifications and benchmarks are not the same... 5.2 GB/sec switch6.4 GB/sec switch ES40 6/667 DS20 6/500 DS10 6/466 1338.5 1323.0 X-BAR Cross-Bar Performance (Memory Bandwidth)
  • Slide 40
  • 3U Rackmount/Desktop Enclosure UNI processor - EV6 466 MHz 2MB L2 Cache 1.3 GB/sec peak Memory Bandwidth 1 GB Memory (2 GB future) Onboard IDE Controller 30 GB Internal Storage (3 x 10 GB IDE drives) 4 PCI slots (3 64-bit PCI, 1 32-bit PCI) Dual embedded 10/100bT Ports Remote Console Management Switch Selectable Power Minimum Revision O/S Support: UNIX 4.0F, OVMS V7.1-2, Linux AlphaServer DS10 Product Description DS10 256 MB 466: $4600 616: $5200 Tru64 UNIX, 128 MB: $5,798 That Compaq box is a fast box; its a screamer... its the fastest single processor system weve ever experienced. Gray Watson, Lycos a neatly packaged, powerful server that we can stack in our racks in the most cost-effective server density. The performance is incredible and the price is right. Jan Isley, MindSpring 5.2 H x 17.5 W x 19 D October Upgrade: EV67/616 motherboard with integrated graphics, Enet, and SCSI adaptor SPECint_rate95 DS10 6/466: 222 DS10 6/616: 317 (43%)
  • Slide 41
  • 1U Rackmount/Desktop Enclosure UNI processor - EV6 466 MHz, 2MB L2 Cache 1 GB Memory 1.3 GB/sec peak Memory Bandwidth 1 64-bit PCI slot Dual embedded 10/100bT Ports Onboard IDE Controller, dual channel 2 Internal IDE drives OR 1 SCSI drive Optional slim CDROM/floppy combo drive Support for external SCSI, Mem Channel, TruClusters Enhanced manageability and availability Three Year Warranty Minimum Revision O/S Support: UNIX 4.0F, OVMS V7.1-2, Linux (Redhat 6.0 & SuSE 6.1) AlphaServer DS10L Product Description DS10L Jan-Mar 2000 1.6 H x 17.5 W x 20.5 D Remote Management Console (RMC) Power status management and component monitoring Remote power on/off, halt, and reset Dialing a pager phone number Auto Reboot (AR) Automatically reboot the operating system after most system failures Power-up diagnostics October Upgrade: EV67/616 motherboard with integrated graphics, Enet, and SCSI adapter 256 MB 466: $4700 616: $5300
  • Slide 42
  • New Compaq AlphaServer DS10/10L 5 10 15 20 25 30 35 40 HP B2000 440 Sun Ultra 10 440 SPECint95 35.3 25.3 31.8 IBM RS/6000 44P-170 400 18.1 SPECweb99 0 10 20 30 40 50 60 Compaq AlphaServer DS10L IBM RS/6000 44P-170 400 HP B2000 440 SPECfp95 56.1 47.9 52.4 22.7 Sun Ultra 10 440 Compaq AlphaServer DS10L Dell Prec420 866 33.6 0 100 200 300 400 500 600 552 ? 460 IBM RS/6000 44P-170 400 Sun Ultra 10 440
  • Slide 43
  • DS10/DS10L/XP900 Product Directions DS10 DS10L XP900 H1 01 PLANS 1-2-01.PPT
  • Slide 44
  • DS10 New Hotswap Storage Cage DS10 Enclosure holds: 1 FDD (floppy) 1 CDROM 1 internal HDD 1 Storage Cage DS10 with Hotswap CageDS10 with Standard Cage Storage Cage INT HDD FDD CDROM DS10 Standard Cage holds: 2 Internal drives OR 1 Tape 1 internal drive INT HDD STANDARD CAGE INT HDD FDD CDROM DS10 INT HDD FDD CDROM DS10 INT HDD OR Hotswap Cage holds: 2 Hotswap drives 1 internal drive INT HDD HOTSWAP CAGE April 2001
  • Slide 45
  • Dual cpus - EV6/500MHz, 4MB L2 cache per cpu Dual 667 MHz, 8 MB cache 4GB Memory (up to 8 GB) 5.2 GB/sec Memory Bandwidth (crossbar peak) 10 bays, 7 hot-swap drive bays 128 GB Internal Storage (Pedestal: seven 3 1/2 18 GB drives) 108 GB Internal Storage (Rack: six 1 or four 1.6 drives) 6 PCI slots (five 64-bit PCI, one 64-bit PCI/ISA) 532 MB/sec I/O throughput (dual 64-bit PCI) 64 bit PCI Ultra2 SCSI RAID controllers, Ultra3-ready StorageWorks 10/100 Fast Ethernet Hot Swap Power and Fans with N + 1 Integrated Remote Management Console 3 Year On-Site Warranty, Worldwide Service Minimum O/S Revision: UNIX 4.0E, OVMS 7.1-2, Linux AlphaServer DS20E Product Description DS20E Pedestal Rack/Desktop: 5U New (256 MB) $12,000 500MHz $15,000 667MHz New
  • Slide 46
  • DS20E Design Compact 8.7(5U) Rack design 17.5 2 Alpha CPUs, Upgradeable 2 hot-plug power supplies, 3rd redundant supply option Hot-plug Ultra2 disk drives -- common with ProLiant -- up to 72.8GB! Up to 4GB ECC SDRAM accessed at 5.2GB/s Six 64-bit PCI slots Hot-swap fans
  • Slide 47
  • 1323 882.4 358 261 242.4 0 200 400 600 800 1000 1200 1400 SGI Origin 2000 IBM RS6000 Sun E 6001 HP C180 AlphaServer DS20E 6/500 Memory Bandwidth in MB/Sec McCalpin Streams TRIAD* DS20E Memory Bandwidth
  • Slide 48
  • 361 494 712 249 602 400 300 200 100 500 CPUS: 1 2 DS20E 600 700 DS20 Alpha 1200 5/533 HP L2000 440 Mhz Alpha 4100 5/600 Alpha DS20 6/500 IBM SP 375 Mhz Alpha DS20E 6/667 SUN Ultra 60 450 Mhz 175 353 169 334 149 291 219 438 4100 1200 System Throughput SPECint_rate95
  • Slide 49
  • AlphaServer DS20E Product Directions
  • Slide 50
  • The Compaq AlphaServer DS20E Series Next Steps 833MHz EV68
  • Slide 51
  • AlphaServer DS20E EV68 Speedup Dual 833MHz Alpha 21264A EV68 based on Tsunami chip set Faster CPU speed and 8MB cache using dual data-rate technology New Ultra3 6-slot universal disk card cage New 64-bit PCI Ultra2 SCSI RAID Ultra3-StorageWorks adapters and disk drives Tru64 UNIX, OpenVMS, and Linux support Available as new system or CPU daughter card upgrade Q2 2001
  • Slide 52
  • AlphaServer ES40 Product Directions
  • Slide 53
  • 1-4 cpus - EV6/500MHz, 4MB L2 cache per cpu 1-4 667 MHz, NEW 833 MHz, 8 MB cache 16 GB, up to 32 GB Memory 5.2 GB/sec Memory Bandwidth (crossbar peak) 6 or 10 64-bit PCI 532 MB/sec I/O throughput (dual 64-bit PCI) PCI RAID subsystems, Ultra2 SCSI 10,000 RPM drives 4 to 66 internal hot swap storage bays 145 GB to 1.2 TB of internal storage 10/100 Fast Ethernet Hot Swap Power and Fans with N + 1 Integrated Remote Management Console 3 Year On-Site Warranty, Worldwide Service Minimum O/S: UNIX 4.0F, OVMS 7.1-2, Linux AlphaServer ES40 ES40 EV67 8 MB cache 32 GB memory 30 - 40% performance increase New 833 MHZ ES40 $32,000 Linux no disk $25K NEW 833 MHz Clock speed is 25% higher Oracle Apps increased 30% Celera has seen 72% !!! Compaqs AlphaServer ES40 is a far better balanced and much more scalable system than any other we have investigated... able to handle tasks that only a year ago would have required a much larger, much more expensive system Marxhall Peterson, VP IT, Celera Genomics
  • Slide 54
  • 5U, 8.75 Same architectural features as the ES40 8U system Same OS versions: OpenVMS, Tru64 UNIX, Linux* AlphaServer ES40LP Low-Profile 5U Rackmount for High Density Computing and Embedded Solutions 7 systems in a 79 H9A15 cabinet MIL-STD-167 (vibration resistant) ES40 ES40LP 64-bit PCI slots107 Integrated NINoYes Internal disk 82 Redundant power supplyYesNo Front-access tape driveYesNo Interfaces2 serial, 1 parallel, 2 USB, Remote Server Mgmt port TP Ethernet port DIFFERENCES: *(custom quote for LP)
  • Slide 55
  • SDRAM Memory Up to 32GB MMB 2 MMB 1 MMB 0 Serial, Parallel keyboard/mouse floppy Cache 8 MB per CPU Compaq AlphaServer ES40 21264 System Architecture 256b 83Mhz (2.6GB/S) 256b 83Mhz (2.6GB/S) Each @ 64b 333MHz (2.6GB/S ) EV6/500 EV67/667 Mhz PCI5PCI4 PCI0 PCI2 PCI1 PCI6PCI7 PCI-USB PCI-junk IO PCI3PCI8 PCI 9 64b 33MHz (266MB/S) MMB 3 Quad C-Chip Controller PCI Chip Bus 0 PCI Chip Bus 1 D D D D DD D D
  • Slide 56
  • ES40 - Electronic Modules MLB DIMMs CPUs Memory Modules PCI Backplane I/O Module
  • Slide 57
  • 85 120 190 260 0 50 100 150 200 250 300 SAP Benchmark - S&D Users Bull PPC 604e 255 MHz Sales and Distribution users IBM RS/6000 F50 SUN UE 450 (300 MHz) Alpha 4100 5/600 SAP R3 Two-Tier Client Server with quad processor systems and Oracle 270 Alpha ES40 6/500 Alpha ES40 67/667 ES40 400 350 400
  • Slide 58
  • Leadership ES40 Commercial Performance Oracle Applications Benchmark 3,528 Audited Results Users
  • Slide 59
  • The Compaq AlphaServer ES40 Series Next Steps - ES40 833MHz EV68
  • Slide 60
  • AlphaServer ES40 EV68 Speedup Quad 833MHz Alpha 21264A EV68 based on Typhoon chip set Faster CPU speed and 8MB cache using dual data-rate technology More capacity with maximum memory to 32GB New Ultra3 6-slot universal disk card cage New 64-bit PCI Ultra3 SCSI RAID coming in Q1 Ultra3-StorageWorks adapters and disk drives Tru64 UNIX, OpenVMS, and Linux support Available as new system or CPU daughter card upgrade Q1 2001
  • Slide 61
  • Alpha Leadership Integer Performance
  • Slide 62
  • Alpha Leadership Technical Performance
  • Slide 63
  • 1525 1232 0 1000 2000 3000 4000 5000 3528 USERS Alpha ES40/667 SUN E450 IBM H70 Oracle Applications quad processors 4563 Alpha ES40/833 30% increase with new ES40!
  • Slide 64
  • 56.8 HP N4000 550 MHz 50 0 25 104 34.3 75 60.1 SUN Ultra 60 Model 1450 450 MHz (512 MB) HP N4000 550 MHz 81.1 SUN Ultra 60 Model 1450 450 MHz (512 MB) 31.9 IBM RS/6000 44P M170 P3-II 450 MHz 57.6 71.5 71.9 ES40 EV6/667 Operations per second HP J5600 552 MHz source: www.spec.org February, 2001 HP C3600 552 MHz ES40 EV6/833 100 ES40 EV6/833 83.5 ES40 EV6/667 SPECjvm98 Java Benchmark 2 cpus 1 GB 1 cpu 1 GB
  • Slide 65
  • 0 50 100 150 IBM S80PL8000ES40/833 38 K ~195 users 46 K ~220 users 115 K ~510 users Dialog (query phase) thousand steps/hour 0 75 150 225 IBM S80PL8000ES40/833 10 Million 168 Million 14 Million Realignment (change run) million rows/hours 4 x EV68 833/8MB 8 GB RAM, BW 2.0B, Oracle 8.i 0 25 50 IBM S80 8 x PIII 700/2MB 4 GB RAM, 420GB DB BW 1.2B, SQL 2000 24 x RS63-III 450/8MB 8 GB RAM, 900GB DB BW 1.2B, DB2 PL8000ES40/833 2 Million 25 Million 3 Million Throughput (load phase) million rows/hour SAP BW Benchmark New 4-way ES40 outperforms 24-way IBM S80! 6 X more cpus for only 2+ more users
  • Slide 66
  • 2001 1GHz 2001 Next Steps - The Compaq AlphaServer ES40 Series Next Steps - ES45
  • Slide 67
  • AlphaServer ES45 Leadership CPU Performance 1GHz EV68 Alpha Processors 8MB High-speed 333Mhz DDR Cache Superior Memory Performance Dual 256bit 125Mhz Memory Bus 8GB/sec Memory bandwidth Enhanced I/O Performance 4x 64-bit PCI Busses (10 slots) 1.85GB/sec I/O Bandwidth Hot-swap PCI (7) AGP 4X (replaces 2 PCI slots) Q3 2001
  • Slide 68
  • Superior Memory Bandwidth TitanC-Chip Address Arrays 0 & 2 Address Arrays 1 & 3 256-bit Data Bus 1 256-bit Data Bus 0 BUS 0 256bit 125Mhz 4.0GB/sec BUS 1 256bit 125Mhz 4.0GB/sec
  • Slide 69
  • Enhanced I/O Performance 3 PCI Channels 64bit/66Mhz 6 slots 1 PCI Channel 64bit/33Mhz 4 slots 1.85GB/sec I/O Bandwidth 5V & 3.3V PCI Options
  • Slide 70
  • Introducing PCI Hot Plug Enhancements 7 Hot-Plug PCI Slots 4 - 64bit/66Mhz 3 - 64bit/33Mhz
  • Slide 71
  • SDRAM Memory Up to 32GB MMB 2 MMB 1 MMB 0 Serial, Parallel keyboard/mouse floppy Cache 8 MB per CPU 256b 83Mhz (2.6GB/S) 256b 83Mhz (2.6GB/S) 256b 125Mhz (4.0GB/S) 256b 125Mhz (4.0GB/S) Each @ 64b 333MHz (2.6GB/S ) EV67/667 EV68/833 Mhz Each @ 64b 500MHz (4.0GB/S) EV68 1.00Ghz PCI5PCI4 PCI0 PCI2 PCI1 PCI6PCI7 PCI-USB PCI-junk IO PCI3PCI8 PCI 9 64b 33MHz (266MB/S) 64b 66MHz (528MB/S ) PCI5PCI4 PCI0 PCI2 PCI1 64b 66MHz (528MB/S) PCI6PCI7 PCI-USB PCI-junk IO PCI3PCI8 PCI 9 64b 33MHz (266MB/S) 64b 66MHz (528MB/S) PCI9 HSPCI8 HSPCI7 HSPCI6 HS PCI3 HSPCI2 HS Quad C-Chip Controller PCI Chip Bus 0 PCI Chip Bus 1 D D D D DD D D PCI1 HS Quad C-Chip Controller PCI Chip Bus 0,1 PCI Chip Bus 2,3 D D D D DD D D MMB 3 PCI Chip Bus 2,3 AGP 4X PCI0 PCI2 PCI1 PCI-USB PCI-junk IO PCI3 64b 33MHz (266MB/S) PCI3 HSPCI2 HS 32b 266MHz (1056MB/S) AlphaServer ES40 System Architecture AlphaServer ES45 System Architecture
  • Slide 72
  • Extending Alpha Leadership in Integer Performance SPECint2000 275 379 442 433 540 663 0 100 200 300 400 500 600 700 800 ES40 1.0 Ghz ES40 833Mhz ES40 667Mhz Intel Pentium III 1.0GHz IBM M80 RS64 500Mhz HP9000 N4000 552Mhz
  • Slide 73
  • SPECfp2000 369 250 335 562 662 832 0 100 200 300 400 500 600 700 800 900 ES40 1.0 Ghz ES40 833Mhz ES40 667Mhz Intel Pentium III 1.0GHz IBM M80 RS64 500Mhz HP9000 N4000 552Mhz Extending Alpha Leadership in Technical Performance
  • Slide 74
  • What is the new GS series? GS160 GS80 GS320 32-way 16-way 8-way 400+ systems shipped Q3 $300M engineering investment 200+ First Day orders D.H. Brown Associates, Inc UNIX Server Pricing and Configuration Monitor, Oct. 13, 1999 ... offers Blazing Speed a system hardware partition designed to support multiple, diverse operating systems extreme operational flexibility via hot add/hot swap components modularity... allows for smooth scalability and capacity growth easy upgrades to future Alpha processors while preserving initial investments much more than just better performance... Q4 vs. Q3: 15% increase in GS system shipments Y2000 GS products/services: $800 Million
  • Slide 75
  • SMP or MPP ? Analysts View Users requiring high-performance UNIX database systems should fully exploit SMP - with its mature software library - before considering MPP systems. Most users will never use up the increasing power of SMP systems. Gartner Group Users requiring high-performance UNIX database systems should fully exploit SMP - with its mature software library - before considering MPP systems. Most users will never use up the increasing power of SMP systems. Gartner Group Almost all commercial applications are currently best served by shared-everything server architectures. META Group Almost all commercial applications are currently best served by shared-everything server architectures. META Group Especially for Single Stream Batch Jobs... One horse is better than 100 chickens. Gartner
  • Slide 76
  • Achieving High Application Performance Fewer, more powerful nodes Higher reliability & availability Easier application porting/tuning 64-bit Very Large Memory capability Less complex system management Easier growth and scalability Automatic dynamic load balancing Lower cost of ownership Better resource utilization Shared Peripherals Best price/performance SMP MPP SMP Advantages
  • Slide 77
  • Alpha High-Bandwidth Memory CPU MEM System elements share a bus Bus bandwidth is shared by all processors and I/O Data must be transferred in relatively small blocks Latency can vary significantly dependent upon number of outstanding transactions NEW Cross-bar topology System elements dynamically connected over multiple paths Dedicated bandwidth for large blocks of data between each pair Simultaneous connection to all CPUs Well-defined latency, limited request queueing CPU MEM Previous Bus-based topology Not Just the Fastest Processor... Switch
  • Slide 78
  • GS80/160/320 Modular Architecture A scalable architecture from $100K to $2M Common architectural components Packaging optimized for office use and maximum expansion Less complexity than numerous discrete products Benefits: Common architecture for deployment Maximum investment protection Existing GSxx AlphaServer Applications New HPTC and Enterprise Applications New Scalable Enterprise Applications GS80GS160GS320 CommonComponents CPUs Memories 2 Backplanes PCI Buses
  • Slide 79
  • Quad Building Block One to four CPUs per building block EV67, 729MHz, upgrades to EV68, EV7 (4) 1.6 GB/sec switched CPU/memory connections simultaneously 6.4 GB/sec total memory bandwidth Up to 4GB per memory board, 16 GB per building block, (8 GB/32 GB total) 32-way interleaving (EV6: 32 in-flight instructions) 1.6 GB I/O Bandwidth, 8 PCI buses; 28 slots No slot trade-offs for CPUs, memory, or I/O Switch EV6 Mem I/O PCI 4 Port Adapter 2GB/sec
  • Slide 80
  • EV6 Mem I/O Switch 4 Port Adapter 4 slots 3 slots 4 slots 3 slots 4 slots 3 slots 4 slots 3 slots 8 64-bit PCIs Rack-mounted I/O Strategy 28 PCI slots available to each building block The bandwidth of 8 64-bit PCI busses for each building block Flexible rack-mount packaging for PCIs Migration strongly encouraged to those that best support I/O demands: KZPBA, KGPSA, KZPAA CIPCA, Memory Channel II DEFPA, DE500, ATM Graphics Design supports slot-intensive AND bandwidth-intensive applications
  • Slide 81
  • EV6 Mem I/O Switch 4 Port Adapter 2GB/sec PCI EV6 Mem Global Port I/O Switch 4 Port Adapter 2GB/sec PCI Scalability With Building Blocks Flexible configurations of CPUs, memory and I/O Low-latency 1.6 GB/s global port: Extremely low local vs remote memory latency (>3:1) SMP programming model Up to 8 processors EV67, EV68 2 x 6.4 GB/s aggregate memory bandwidth 32 GB of memory Up to 3.2 GB/sec I/O bandwidth, 56 PCI slots Addition, not shared division
  • Slide 82
  • EV6 Interconnect EV6 Interconnect 5-8 CPUs Optional 1-4 CPU Box EV6 Mem I/O Switch EV6 Mem I/O Switch Entry System Package 1-4 CPUs in basic rack-mount box configuration 5-8 CPU expansion in optional 2nd rack-mount box Up to 64 GB of memory Up to sixteen 64-bit PCI buses, 56 PCI slots Single phase power, office standards Two Quad Building Blocks GS80
  • Slide 83
  • 12.8 GB/sec global switch (1.6 GB per connection) Up to 32 CPUs with 256 GB of memory in 3-bay cabinet CPU-Memory bandwidth of 51.2 GB (6.4 per QBB) Up to sixty-four 64-bit PCI buses; 224 slots; 12.8 GB I/O bandwidth NEW!NEW! building blocks can support different speed cpus EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV67 EV68 Global Switch GS320 A 32 Processor SMP AlphaServer extensive experience with high performance switches VAX9000, GIGAswitch, MC2, DS/ES systems non-blocking architecture memory references do not interfere with each other low latency even under heavy loads new low occupancy cache coherency protocol 20 patents filed for this hierarchical switch alone over 50 patents for entire system MTBF of 10-15 years
  • Slide 84
  • Front ViewRear View 1 - 4 Quad Bldg Blocks Global Switch Capacity on Demand 5 - 8 Quad Bldg Blocks I/O & Power GS320 1. Pre-install system boxes, QBBs, CPUs and memory; call to enable; no operational disruption or delays 2. Pre-install system boxes and QBBs only; hot add CPUs and memory; take advantage of faster cpus, denser memory
  • Slide 85
  • CPU Memory MTBF of microprocessor is very high due to VLSI content (millions of hours) CPU module raw MTBF = 291,300 hours (>33 years @ 24x7 operation) Availability improvement due to cache ECC and CPU re-try = ~50% Net CPU failure rate = 1 failure per 50 years of run time CPU failure does not necessarily cause a system outage CPUs are hot swappable 1 GB memory module raw MTBF = 94,800 hours (>11 years @ 24x7 operation) Availability improvement due to ECC = ~100 - 200% Net memory failure rate = 1 failure per 22 - 44 years of run time Memory failure does not necessarily cause a system outage Bad memory can be mapped-out Base Component Reliability Profile
  • Slide 86
  • Distributed power regulation N+1 power (per cabinet) Hot-swap and hot-add Dual-AC power source option Feed A Feed B GS160/320 Power Redundancy Automatic and transparent failover should one power source fail
  • Slide 87
  • Hierarchical Switch Availability Profile High-reliability architecture and design ECC: 4 out of 5 errors predicted to be transient, not hard Large ASICs yield 10-15 years MTBF N+1 power regulators Redundant cooling Designed for ease of replacement: ~ 45 minutes Individual QBBs continue to operate
  • Slide 88
  • Clock System Availability Distributed clocking limits critical central resources Central clock on H-switch Clock splitters Central clock A potential single point of failure Demonstrated reliability of 1 million hours Easily accessible for replacement
  • Slide 89
  • Redundant, High-Function Console Console bus (red) connects ~100 microprocessors Full system visibility Full system control Full event logging Redundant console connections Automatic failover Up to 8 connections Redundant System Management Consoles Up to 8 supported Dial-in/Dial-out Global Switch EV67 Mem I/O Switch EV67 Mem I/O Switch EV67 Mem I/O Switch EV67 Mem I/O Switch EV67 Mem I/O Switch EV67 Mem I/O Switch EV67 Mem I/O Switch EV67 Mem I/O Switch SCM PCI PCI OCP LAT PCI PCI PCI PCI PCI PCI
  • Slide 90
  • Reliable and Redundant I/O Subsystem Multiple buses within each PCI drawer Multiple drawers per building block Hot-swap N+1 power in each PCI drawer Low cost of redundancy, including master PCI Quad Building Block PCI Drawer 4 PCI Buses in 2 Segments 2nd PCI Drawer
  • Slide 91
  • Tru64 UNIX: I/O Optimizations I/O Multi-pathing Multiple concurrent paths to storage (up to 64) Transparent failover Adaptive load balancing I/O adapter affinity: preference given to closest adapter Benefits No single point of failure configurations Supports online repair & upgrades Reduces host interconnect hot-spots I/O affinity yields high I/O scalability for large numbers of processors FC Switch HSG80 EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch Global Switch PCI Subsystem Cabinets can be up to 10 meters from servers With fibre, up to 3 Km For Disaster Tolerant, 10 Km
  • Slide 92
  • Global Switch Modular GS Series Architecture movement of resources from one partition to another without affecting application availability no reboot, no reconfiguration Hardware Partitions Software Partitions (subdivision of main partition) more granular control of resources for workload management guarantee cpu/memory for specific applications dynamic reallocation without reboot or reconfiguration L L F F E E I I X X T T L L B B Y Y I I I I EV6 Mem Switch I/O EV6 Mem Switch I/O EV6 Mem Switch I/O EV6 Mem Switch I/O EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch EV6 Mem I/O Switch Dynamic reconfig of partitions by QBB OpenVMS Galaxy, Tru64 UNIX (5.1x) Different/Mixed application workloads Different/Mixed speed cpus, memory, I/O Different/Mixed versions of operating systems (5.0, 5.1 or VMS & UNIX) Hardware firewalls between partitions for software isolation & fault containment EV67 EV68 5.0 5.1 UNIX VMS
  • Slide 93
  • Server Consolidation Replaced with (2) GS160 (8 CPU -- partitioned, 32GB) with FibreChannel storage U.S. Government Agency App run time reduced from 24 hours to 6. One task reduced from 1.5 hours to 20 minutes. Customer believes they can reduce run times by another third! Sorts running 6x faster Still room for even more optimizations! Customer started with (2) 8400 (6 CPU, 4GB), (2) 2100, (6) AS600 with HSZ-based storage
  • Slide 94
  • SAP Performance SD Benchmark Users Two Tier R/3 V4.0B 1000 2000 3000 1100 400 Wildfire: 88% scaling ES40 EV67/667 4 cpus 2720 GS320 EV67/729 32 cpus AlphaServer GS320 (Wildfire) 32 cpus IBM iSeries 400 24 cpus Wildfire delivers 2nd best performance with 1/2 the cpus 3000 Fujitsu Siemens GP7000F 64 cpus Fujitsu Siemens GP7000F Model 2000 UltraSparc 300 MHz 64 cpus 1410 SUN UE 10000 64 cpus Informix source: ideasinternational.com 1708 IBM RS/6000 S80 DB2 Oracle
  • Slide 95
  • 0 25 50 75 100 125 IBM S80GS320 Million rows/hour 8 x PIII 700/2MB 4 GB RAM, 420GB DB BW 1.2B, SQL 2000 24 x RS63-III 450/8MB 8 GB RAM, 900GB DB BW 1.2B, DB2 PL8000ES40/833 2 Million 114 Million 25 Million 3 Million Throughput (load phase) SAP BW Benchmark 0 50 100 150 200 250 IBM S80GS320 Thousand steps/hour PL8000ES40/833 38 K ~195 users 207 K ~1000 users 46 K ~220 users 115 K ~510 users Dialog (query phase) 0 75 150 225 300 375 IBM S80GS320 Million rows/hour 32 x EV67 731/4MB 32 GB RAM, 1.3TB DB BW 2.0B, Oracle 8.i PL8000ES40/833 4 x EV68 833/8MB 8 GB RAM, BW 2.0B, Oracle 8.i 10 Million 313 Million 168 Million 14 Million Realignment (change run) 2 X more users 1/3 more cpus
  • Slide 96
  • 25 X from 733 MHz to 1.5 GHz from 4 to 8 instructions per cycle from 2 to 10 GB/sec memory bandwidth simultaneous multi-threading (4 threads) from 32 to 48 cpus in single cab Performance Improvements 2 X 1.5 X 2000 EV67 2004 EV8 its more than MHz What matters at the end of the day, megahertz or getting your work done? Terry Shannon, ZDnet News, May 29, 2000 real application performance