Post on 21-Jan-2016
SupercomputersSupercomputersNew PerspectiveNew Perspective
Prof. Sin-Min LeeProf. Sin-Min Lee
Department of Computer Department of Computer ScienceScience
Multiple Processor Multiple Processor OrganizationOrganization
Single instruction, single data stream Single instruction, single data stream - SISD- SISD
Single instruction, multiple data Single instruction, multiple data stream - SIMDstream - SIMD
Multiple instruction, single data Multiple instruction, single data stream - MISDstream - MISD
Multiple instruction, multiple data Multiple instruction, multiple data stream- MIMDstream- MIMD
Single Instruction, Single Single Instruction, Single Data Stream - SISDData Stream - SISD
Single processorSingle processor Single instruction streamSingle instruction stream Data stored in single memoryData stored in single memory Uni-processorUni-processor
Single Instruction, Multiple Single Instruction, Multiple Data Stream - SIMDData Stream - SIMD
Single machine instruction Single machine instruction Controls simultaneous executionControls simultaneous execution Number of processing elementsNumber of processing elements Lockstep basisLockstep basis Each processing element has associated Each processing element has associated
data memorydata memory Each instruction executed on different set Each instruction executed on different set
of data by different processorsof data by different processors Vector and array processorsVector and array processors
Multiple Instruction, Single Multiple Instruction, Single Data Stream - MISDData Stream - MISD
Sequence of dataSequence of data Transmitted to set of processorsTransmitted to set of processors Each processor executes different Each processor executes different
instruction sequenceinstruction sequence Never been implementedNever been implemented
Multiple Instruction, Multiple Multiple Instruction, Multiple Data Stream- MIMDData Stream- MIMD
Set of processorsSet of processors Simultaneously execute different Simultaneously execute different
instruction sequencesinstruction sequences Different sets of dataDifferent sets of data SMPs, clusters and NUMA systemsSMPs, clusters and NUMA systems
Taxonomy of Parallel Taxonomy of Parallel Processor ArchitecturesProcessor Architectures
MIMD - OverviewMIMD - Overview
General purpose processorsGeneral purpose processors Each can process all instructions Each can process all instructions
necessarynecessary Further classified by method of Further classified by method of
processor communicationprocessor communication
Tightly Coupled - SMPTightly Coupled - SMP
Processors share memoryProcessors share memory Communicate via that shared memoryCommunicate via that shared memory Symmetric Multiprocessor (SMP)Symmetric Multiprocessor (SMP)
Share single memory or poolShare single memory or pool Shared bus to access memoryShared bus to access memory Memory access time to given area of Memory access time to given area of
memory is approximately the same for memory is approximately the same for each processoreach processor
Tightly Coupled - NUMATightly Coupled - NUMA
Nonuniform memory accessNonuniform memory access Access times to different regions of Access times to different regions of
memroy may differmemroy may differ
Loosely Coupled - ClustersLoosely Coupled - Clusters
Collection of independent Collection of independent uniprocessors or SMPsuniprocessors or SMPs
Interconnected to form a clusterInterconnected to form a cluster Communication via fixed path or Communication via fixed path or
network connectionsnetwork connections
Parallel Organizations - SISDParallel Organizations - SISD
Parallel Organizations - Parallel Organizations - SIMDSIMD
Parallel Organizations - Parallel Organizations - MIMD Shared MemoryMIMD Shared Memory
Parallel Organizations - Parallel Organizations - MIMDMIMD
Distributed MemoryDistributed Memory
Symmetric MultiprocessorsSymmetric Multiprocessors with the following characteristicswith the following characteristics
Two or more similar processors of comparable capacityTwo or more similar processors of comparable capacity Processors share same memory and I/OProcessors share same memory and I/O Processors are connected by a bus or other internal connectionProcessors are connected by a bus or other internal connection Memory access time is approximately the same for each Memory access time is approximately the same for each
processorprocessor All processors share access to I/OAll processors share access to I/O
Either through same channels or different channels giving Either through same channels or different channels giving paths to same devicespaths to same devices
All processors can perform the same functions (hence All processors can perform the same functions (hence symmetric)symmetric)
System controlled by integrated operating systemSystem controlled by integrated operating system providing interaction between processors providing interaction between processors Interaction at job, task, file and data element levelsInteraction at job, task, file and data element levels
A stand alone computerA stand alone computer
SMP AdvantagesSMP Advantages
PerformancePerformance If some work can be done in parallelIf some work can be done in parallel
AvailabilityAvailability Since all processors can perform the same functions, Since all processors can perform the same functions,
failure of a single processor does not halt the systemfailure of a single processor does not halt the system Incremental growthIncremental growth
User can enhance performance by adding additional User can enhance performance by adding additional processorsprocessors
ScalingScaling Vendors can offer range of products based on Vendors can offer range of products based on
number of processorsnumber of processors
Block Diagram of Tightly Block Diagram of Tightly Coupled MultiprocessorCoupled Multiprocessor
HistoryHistory Cray Research founded in 1972.Cray Research founded in 1972. Cray Computer founded in 1988.Cray Computer founded in 1988. 1976 First product – Cray-1 (240,000,000 OpS). 1976 First product – Cray-1 (240,000,000 OpS).
Seymour Cray personally invented vector register Seymour Cray personally invented vector register technology.technology.
1985 Cray-2 (1,200,000,000 OpS, a 5-fold 1985 Cray-2 (1,200,000,000 OpS, a 5-fold increase from Cray 1). Seymour is credited with increase from Cray 1). Seymour is credited with immersion-cooling technologyimmersion-cooling technology
Cray-3 used revolutionary new gallium arsenide Cray-3 used revolutionary new gallium arsenide integrated circuits for the traditional silicon onesintegrated circuits for the traditional silicon ones
1996 Cray was bought by SGI1996 Cray was bought by SGI In March 2000 the Cray Research name and In March 2000 the Cray Research name and
business was sold by SGI to Tera Inc.business was sold by SGI to Tera Inc.
Visual TourVisual Tour
Market Segment (from top500)Market Segment (from top500)
Count Share % Rmax Sum (GF) Rpeak Sum (GF) Procs Sum
Pentium4 Xeon 232 46.4 422375 810977 142065
Itanium2 84 16.8 190357 243070 42048
HP 50 10 52503 92016 26064
Opteron 30 6 61729 102444 24796
Power4+ 26 5.2 90730 174387 24042
Power4 19 3.8 34874 68089 12966
Alpha 12 2.4 32414 46182 23512
Nec 10 2 47177 52669 6488
PowerPC 8 1.6 127194 176979 51664
Power 8 1.6 24465 36742 27732
Cray 7 1.4 16298 17869 1396
Hitachi SR8000 4 0.8 5152 6314 1548
Sparc 4 0.8 12076 24535 5348
MIPS 2 0.4 2461 4301 7168
Pentium4 1 0.2 1212 5753 940
Intel 1 0.2 2379 3207 9632
AMD Athlon 1 0.2 1017 1400 500
Power5 1 0.2 3068 5472 720
All 500 100 % 1127481 1872406 408629
Supercomputer Architecture Supercomputer Architecture
Count Share % Rmax Sum (GF) Rpeak Sum (GF) Procs Sum
Scalar 479 95.8 1058853 1795555 399197
Vector 21 4.2 68628 76851 9432
All 500 100 % 1127481 1872406 408629
Current Cray ProductsCurrent Cray Products
Cray X1 is the only Cray’s product with Cray X1 is the only Cray’s product with a unique vector CPUa unique vector CPU Competitors are: Fujitsu, NEC, HPCompetitors are: Fujitsu, NEC, HP
Cray XT3 and XD1 use AMD Opteron Cray XT3 and XD1 use AMD Opteron CPUs (series 100 and series 200 CPUs (series 100 and series 200 accordingly)accordingly)
You can find full product specifications You can find full product specifications as well as additional information on as well as additional information on current systems at www.cray.comcurrent systems at www.cray.com
Performance MeasurementsPerformance Measurements Performance is measured in teraflopsPerformance is measured in teraflops Linpack is a standard benchmarkLinpack is a standard benchmark Performance is also measured in memory Performance is also measured in memory
bandwidth & latency, disk performance, bandwidth & latency, disk performance, interconnects, internal IO, reliability, and interconnects, internal IO, reliability, and othersothers
For example:For example: My home system, Athlon 750, gives about 34 My home system, Athlon 750, gives about 34
megaflops (34*10^6 flops)megaflops (34*10^6 flops) Current mid-range supercomputers give about Current mid-range supercomputers give about
40 teraflops(40*10^12 flops) which is 1,176,470 40 teraflops(40*10^12 flops) which is 1,176,470 times fastertimes faster
Scalable Architecture in XT-Scalable Architecture in XT-33
Is Cray a good deal?Is Cray a good deal?
Typical Cost approximately $30 million and Typical Cost approximately $30 million and aboveabove
Useful lifetime – 6 yearsUseful lifetime – 6 years Most customers use supercomputers at 90% Most customers use supercomputers at 90%
- 98% load- 98% load Clustered supercomputers and machines Clustered supercomputers and machines
build around common desktop components build around common desktop components (AMD/Intel CPUs, memory chips, (AMD/Intel CPUs, memory chips, motherboards, and etc.) are significantly motherboards, and etc.) are significantly cheapercheaper
FutureFuture Cray’s “Red Storm” System in Sandia Cray’s “Red Storm” System in Sandia
National Laboratories is running on Linux OSNational Laboratories is running on Linux OS Current Cost $90 millionCurrent Cost $90 million Uses 11,648 AMD Opteron CPUsUses 11,648 AMD Opteron CPUs Current operational speed – 41.5 teraflopsCurrent operational speed – 41.5 teraflops Uses unique SeaStar chip, which passes Uses unique SeaStar chip, which passes
messages between thousands of CPUsmessages between thousands of CPUs Upgrades are scheduled to be completed by Upgrades are scheduled to be completed by
the end of 2005 using dual-core Opteronthe end of 2005 using dual-core Opteron Expected to reach 100 teraflops by the end Expected to reach 100 teraflops by the end
of 2005of 2005
Computers and Weather
forecasting
Weather forecasting
Menu
Weather forecasting
Explanation
Overview picture
Data collection
Sensors
Data logging
The Grid System
Weather station
Radiosonde
Satellites
Radar
Weather ships
Supercomputers
Parallel Processing
Software
Pressing Weather forecasting on any slide will bring you back to this menu
During the last two decades the Met Office hasused state-of-the-art supercomputersfor numerical weather prediction and morerecently, also for predictions of global climate.
Weather forecasting
This is a picture of a supercomputer
WeatherWeather forecasting forecasting
Weather forecasters are helped by several things. These include:-
The computer makes millions of calculations.
1. The computer’s advice2. Information from radar
3. Information from satellite pictures
The sums are called differential equations
Before the computer can do the calculations, data has to be collected first.
Weather forecasting
Collecting data on the weather is very important.
Without the data, the computer could not do the calculations that enable it tomake weather predictions.
The next slide shows where the data comes from.
It also shows where the forecasts are sent.
Always remember that the forecasters are highlytrained people and they use their judgementand expertise to make their forecastbased on the information the computergives them and the information from the radarand the satellite pictures.
Supercomputer
Radio-sonde
Weather shipand weather buoys
Satellites
Radar
Weather forecasting
Weatherstations
Forecaster
Supercomputer
Global forecast services
Aviation
ShippingNational and International
Forecast Services up to 7 days ahead.
Radio-sonde
Weather shipand weather buoys
Satellites
Radar
Weather forecasting
Weatherstations
Forecaster
Supercomputer
Global forecast services
Aviation
ShippingNational and International
Forecast Services up to 7 days ahead.
Radio-sonde
Weather shipand weather buoys
Satellites
Radar
Weather forecasting
Weatherstations
Weather forecasting
Data is collected continuously for the computer from the following:-
1. Weather stations2. Automated weather
stations 3. Satellites 4. Radar 5. Radiosondes 6. Weather ships 7. Mini-radiosondes 8. Radar 9. Aeroplanes10. Drifting buoys
Weather forecasting
The data measurements are made by sensors
A sensor is a transducer which responds to some physical property such as pressure, temperature, rate of flow.
A transducer is an electronic component which converts energy from one form to another.
We want the transducers to send signals to the computer in the Met. Office.
1. Temperature .. Air, surface and subsurface temp.
2. Atmospheric Pressure
3. Wind speed
4. Wind direction
5. Humidity
6. Rainfall
Weather forecasting
7. Sunshine
The measurements needed include:-
Weather forecasting
These are some of the sensors used to collect data.
Weather forecasting
Data logging is the capture and storage of data for future use.
All the measurements from the sensors are stored because:-
So data logging is used in weather forecasting.
• The computer processes the data in batches
• People need to refer back to weather data for many reasons
Weather forecasting
The next slides explain more about data collection.
To jump to the details of the computers, click on the computer.
Mountain effects
Formation of clouds
Formation of rain and snow
FrictionRadiation from the
earth
Radiation from the atmosphere
Radiation from the sun
Evaporation and
heat exchange
The atmosphere is split up into a 3-D grid.Each land based grid is about 60km.
Sea
Weather forecasting
The vertical grid you just saw only had 5 levels.
The Met Office uses far more, some computer models work on 40 vertical levels.
Weather forecasting
For UK forecasts the horizontal grid is 15km.
For the global forecasts the horizontal grid is 60km
Sea
We need to measure pressure, temperature,
wind speed and wind directions as well as rainfall, cloud cover,etc in as many grid spaces as possible
Weather forecasting
Sea
Weather forecasting
Weather stations are used here.
A weather station sends signals back to the Met Office computer.
The instruments measure pressure, temperature and humidity.
Weather forecasting
Some weather stations are automated. These send their measurements back to the computerdirectly.
Picture thanks to Sparholt school who have a great web sitewww.hants.gov.uk/sparsholtschoolscentre
We need to measure pressure, temperature,
wind speed and wind directions as well as rainfall, cloud cover,etcRadiosonde
s are used up here.
20 km
Sea
Weather forecasting
A radiosonde sends signals backto the Met Office computer. They are attached to a balloon carrying instruments.
The instruments measure pressure, temperature and humidity.
By measuring the track of the radiosonde, the wind direction and speed can be calculated.
Weather forecasting
Photo courtesy of the British Atmospheric Data Centre
All over the world, radiosondes are released at midday and midnight UTC.
UTC = Universal Time Co-ordinated
Weather forecasting
UTC = Universal Time Co-ordinated
Hawaii 2 amAtlantic Daylight
9 am
Baghdad3 pm
West Australian Standard 7 pm
UTC is also known as GMTGreenwich Mean Time
Here are some of the times when the radiosondesare released to correspond with the ones in Britain at midday.
Weather forecasting
Sea
We need to measure pressure, temperature,
wind speed and wind directions as well as rainfall, cloud cover,etc
Minisondes are used
here.
5 km
Everest is 8.85 Kms high. So we have shown you a very high mountain!
Weather forecasting
A minisonde system takes measurements at 5Km above sea level.
Both radiosondes and minisondes use radio waves to transmit the data readings to the computers
Weather forecasting
Sea
We need to measure pressure, temperature,
wind speed and wind directions as well as rainfall, cloud cover,etc
Aeroplanes send
data too. 10 km
Weather forecasting
Sea
We need to measure pressure, temperature,
wind speed and wind directions as well as rainfall, cloud cover,etc
Satellites send
data too.
36,000km
Weather forecasting
The Geostationary satellites are 36,000 Km above earth.
There are two types of satellites.
• Geostationary. These stay in the same spot. They orbit the earth at exactly the same speed as the earth rotates. They are very high above earth - 36,000 km.5 geostationary satellites are enough to give
global coverage.
Weather forecasting
• Polar orbiting. These orbit the earth about 14 times a day. They orbit at 1000 km above the earth.
Weather forecasting
A Geostationary satellite stays in the same place with respect to
earth all the time
A Polar orbiting satellite moves above the earth’s surface
Weather forecasting
Picture courtesy of the European Organisation for the Exploitation of Meteorological Satellites and photo
courtesy of Dundee University.
The URL ( Uniform Resource Locator ) www.sat.dundee.ac.uk/tour.html will tell
you more about satellites if you are interested
Meteosat: a geostationary satellite. It sends microwave signals back to earth.
The satellite’s signals are received by antennae.
©Eumetsat 2003
This is a Polar Orbiting Satellite
Weather forecasting
The satellite can take readings across the entire earth during the course of one day.
Sea
Weather forecasting
Radar systems are used here.
Radar stands for radio detection and ranging. Radio waves are transmitted, when they hit a rain cloud they bounce
back to earth and measurements can be taken.
Weather forecasting
There is a network of radar systems covering Britain.
This is a radar
station.
The computer and the forecaster receive data from the radar systems.
Radar measurements let the forecaster and the computer know if the radar has seen rain clouds.
Sea
Weather forecasting
There are weather ships out to sea
There are also buoys taking weathermeasurements in the sea
This is a Weather ship.
Weather forecasting
Buoys are used at sea more than weather ships these days. They send their data automatically back to the computer.
Thanks to www.gdfcartophily.co.uk
Weather forecasting
The Meteorological Office uses supercomputers. These are the largest computers in the world.
There will be a new supercomputer in the Met Office’s new location at Exeter.
The Met Office has re-located (moved) to Exeter from Bracknell.Bracknell is near London. Exeter is in the South Westof England.
• Bracknell
• Exeter
Weather forecasting
We store Gigabytes (one thousand million bytes 109) of data on the Hard drives in our school.Supercomputers have stores for Terabytes of data.
A Terabyte isone million, million bytes, 1012 bytes.
A supercomputer is a very large computer, which works very, very fast. It is about 1000 times more powerful than a PC.
It also has an enormous store (memory).
Weather forecasting
A supercomputer is a very large computer, which works very, very fast. It is about 1000 times more powerful than a PC
The Met Office supercomputers areMassively Parallel Processor machines.
It usually has many processors working in parallel to achieve the fast processing time.
Massively Parallel Processor machines use Parallel processing
Parallel processing is the simultaneous use of several processors to perform a single job. A job may be split into a number of tasks eachof which may be processed by any availableprocessor.
Weather forecasting
Weather forecasting
Parallel processing splits jobs up and gives different processors tasks. These all have
to be brought together.
When a computer usesseveral processors in parallel
it is known asparallel processing.
Parallel processing is difficult to program
Weather forecasting
How does parallel processing work?.
287 14 21
1*7
A
2*7
B
4*7 3*7
C D
Let’s give four processors A, B, C, D a sum each
Weather forecasting
If processor A did all four sums it would take longer.
Parallel processing is very complicatedto program but it does make it faster to do loads and loads of sums.
Massively Parallel Processor machines.
MPP systems use a distributed hierarchy of memory. This just means that they have to have a system of accessing the memory available.
MPP systems rely on very high bandwidth communications to move data between memory and between different processors so that they are all kept busy during program execution.
Weather forecasting
ReferencesReferences http://research.microsoft.com/users/gbell/http://research.microsoft.com/users/gbell/
craytalk/sld066.htmcraytalk/sld066.htm http://inventors.about.com/library/inventors/http://inventors.about.com/library/inventors/
blsupercomputer.htmblsupercomputer.htm http://americanhistory.si.edu/csr/comphist/http://americanhistory.si.edu/csr/comphist/
cray.htmcray.htm http://web.mit.edu/invent/iow/cray.htmlhttp://web.mit.edu/invent/iow/cray.html www.top500.orgwww.top500.org http://www.spikynorman.dsl.pipex.com/http://www.spikynorman.dsl.pipex.com/
CrayWWWStuff/CrayWWWStuff/ http://news.zdnet.co.uk/hardware/emergingtech/http://news.zdnet.co.uk/hardware/emergingtech/
0,39020357,39162182,00.htm0,39020357,39162182,00.htm