Supercomputers

Overview of Supercomputers

Presented by:Mehmet Demir

20090694ENG-102

Table of Contents Introduction What are They Used For How Do They Differ From a Personal Computer? Where Are They Now Main Parts of Supercomputers Processor Types Conclusion References

Supercomputers

The category of computers that includes the fastest and most powerful (most expensive) ones available at any given time.

Designed to solve complex mathematical equations and computational problems very quickly.

What are They Used For Climate prediction & Weather forecasting

What are They Used For (cont.)

Computational chemistry Crash analysis Cryptography Nuclear simulation Structural analysis

How Do They Differ From a Personal Computer Cost

range from $100,000s to $1,000,000s Environment

most require environmentally controlled rooms Peripherals

lack sound cards, graphic boards, keyboards, etc. accessed via workstation or PC

Programming language FORTRAN

History

Seymour Cray (1925-1996) Developed CDC 1604 – first fully transistorized

supercomputer (1958) CDC 6600 (1965), 9 MFlops Founded Cray Research in 1972 (now Cray Inc.)

CRAY-1 (1976), $8.8 million, 160 MFlops CRAY-2 (1985) CRAY-3 (1989)

Early Timeline of SupercomputersPeriod Supercomputer Peak speed Location

1943-1944 Colossus 5000 characters per second Bletchley Park, England

1945-1950 Manchester Mark I 500 instructions per second University of Manchester, England

1950-1955 MIT Whirlwind20 KIPS (CRT memory)40 KIPS (Core)

Massachusetts Institute of Technology, Cambridge, MA

1956-1958 IBM 70440 KIPS12 kiloflops

1958-1959 IBM 70940 KIPS12 kiloflops

1959-1960 IBM 7090 210 kiloflops U.S. Air Force BMEWS (RADC), Rome, NY

1960-1961 LARC 500 kiloflops (2 CPUs) Lawrence Livermore Laboratory, California

1961-1964 IBM 7030 "Stretch"1.2 MIPS~600 kiloflops

Los Alamos National Laboratory, New Mexico

1965-1969 CDC 660010 MIPS3 megaflops

Lawrence Livermore Laboratory, California

1969-1975 CDC 7600 36 megaflops Lawrence Livermore Laboratory, California

1974-1975 CDC Star-100100 megaflops (vector),~2 megaflops (scalar)

Lawrence Livermore Laboratory, California

1975-1983 Cray-180 megaflops (vector),72 megaflops (scalar)

Los Alamos National Laboratory, New Mexico (1976)

1975-1982 ILLIAC IV150 megaflops,<100 megaflops (average)

NASA Ames Research Center, CaliforniaHad serious reliability problems.

1981-1983 CDC Cyber-205400 megaflops (vector),average much lower.

1983-1985 Cray X-MP 500 megaflops (4 CPUs) Los Alamos National Laboratory, New Mexico

1985-1990 Cray-21.95 gigaflops (4 CPUs)3.9 gigaflops (8 CPUs)

Lawrence Livermore Laboratory and NASALawrence Berkeley National Laboratory (the only 8 CPU system)

1989-1990 ETA-10G10.3 gigaflops (vector) (8 CPUs),average much lower.

1990-1995 Fujitsu Numerical Wind Tunnel

236 gigaflops National Aerospace Lab

1995-2000 Intel ASCI Red 2.15 teraflops Sandia National Laboratories, New Mexico

2000-2002IBM ASCI White, SP Power3 375 MHz

7.226 teraflops Lawrence Livermore Laboratory, California

2002-2004 Earth Simulator 35.86 teraflops Yokohama Institute for Earth Sciences, Japan

2004-Blue Gene/L prototype

70.72 teraflops✝ IBM, Rochester, Minnesota[2]

Where Are They Now

www.top500.org List released twice a year Scores based on Linpack benchmark Solve dense system of linear equations Speed measured in floating point operations

per second (FLOPS)

Architectures - SMP Symmetric Shared-

Memory Multiprocessing (SMP) Share memory Common OS Programs are divided

into subtasks (threads) among all processors (multithreading)

Architectures – MPP Massively Parallel Processing (MPP)

Individual memory for each processor Individual OS’s Messaging interface for communication 200+ processors can work on same application

1. A large retailer wants to know how many camcorders the company sold in

1998, and sends that query to the MPP system. 2. The query goes out to one of the processors which acts as the

coordinator, it breaks up the query for optimum performance. For example, it could break the query up by month; this “sub-query”

then goes to all the processors at the same time.

3. Each sub-query is assigned to a specific processor in the system. To allow this to happen, the database was previously partitioned. For example, a sales tracking database might be broken down by month, and

each processor holds data for one month’s worth of sales information. 4. The responses to the queries are returned to a processor to be coordinated—for

example, each month is added up 5. Final answer is returned to the user.

Architectures – Clustering

Grid computing Many servers connected together Relies heavily on network speed Easily upgraded with addition of more

servers

Processor Types

Vector processing Expensive NEC Earth Simulator

Scalar processing Grid computing

Based on off the shelf parts (ordinary CPUs)

BlueGene/L

IBM MPP (massively parallel processing) #1 on top500 as of November 2004 32,768 processors (700Mhz) 70.72 Teraflops (trillions of FLOPS) Runs linux DNA, climate simulation, financial risk Cost more than $100 million

BlueGene/L System Layout

2 Processors Node communication Mathematical calculations

BlueGene/L Compute Card

BlueGene/L Node Board

BlueGene/L Cabinet

Some of the Others

#2 - Columbia (NASA, USA) – 51.87 TFlops #3 - Earth Simulator (Japan) – 35.86 TFlops #4 - MareNostrum (Spain) – 20.53 TFlops #5 - Thunder (USA) – 19.94 TFlops

The Future

References http://www.top500.org/ http://www.pcquest.com/content/Supercomputer/

102051004.asp http://news.com.com/2100-1008_3-1000421.html?

tag=fd_lede2_hed http://www.research.ibm.com/bluegene/index.html http://www.llnl.gov/asci/platforms/bluegene/papers/

2hardware_overview.pdf http://www.hpce.nec.com/451+M5f7cd421b8e.0.html http://www.cray.com/about_cray/history.html http://www.serc.iisc.ernet.in/~govind/243/L7-PA-Intro.pdf http://www.computerworld.com/hardwaretopics/

hardware/server/story/0,10801,43504,00.html