5PERFORMANCE TRENDSTraditional mainframe/supercomputer performance 25% increase per yearBut … microprocessor performance 50% increase per year since mid 80’s.
6Moore’s Law Moore is co-founder of Intel. 60 % increase per year “Transistor density doubles every 18 months”Moore is co-founder of Intel.60 % increase per yearExponential growthPC costs decline.PCs are building bricks of all future systems.
8Bit Level Parallelism (upto mid 80’s) 4 bit microprocessors replaced by 8 bit, 16 bit, 32 bit etc.doubling the width of the datapath reduces the number of cycles required to perform a full 32-bit operationmid 80’s reap benefits of this kind of parallelism (full 32-bit word operations combined with the use of caches)
9Instruction Level Parallelism (mid 80’s to mid 90’s) Basic steps in instruction processing (instruction decode, integer arithmetic, address calculations, could be performed in a single cycle)Pipelined instruction processingReduced instruction set (RISC)Superscalar executionBranch prediction
10Thread/Process Level Parallelism (mid 90’s to present) On average control transfers occur roughly once in five instructions, so exploiting instruction level parallelism at a larger scale is not possibleUse multiple independent “threads” or processesConcurrently running threads, processes
11Sequential vs Parallel Processing physical limits reachedeasy to programexpensive supercomputers“raw” power unlimitedmore memory, multiple cachemade up of COTS, so cheapdifficult to program
12Amdahl’s LawThe serial percentage of a program is fixed. So speed-up obtained by employing parallel processing is bounded.Lead to pessimism in in the parallel processing community and prevented development of parallel machines for a long time.1Speedup =1-ss +PIn the limit:Spedup = 1/ss
13Gustafson’s LawSerial percentage is dependent on the number of processors/input.Broke/disproved Amdahl’s law.Demonstrated achieving more than 1000 fold speedup using 1024 processors.Justified parallel processing
14Hillis’ Thesis ‘85 Piece of silicon Sequential computer Parallel computerproposed “The Connection Machine” with massive number of processors each with small memory operating in SIMD mode.CM-1, CM-2 machines from Thinking Machines Corporation (TMC)were examples of this architecture with 32K-128K processors. Unfortunately, TMC went out of business.
15Grand Challenge Applications Important scientific & engineering problems identified by U.S. High Performance Computing & Communications Program (’92)
16Flynn’s Taxonomy classifies computer architectures according to: Number of instruction streams it can process at a timeNumber of data elements on which it can operate simultaneouslyData StreamsSingle MultipleSingleSISDSIMDInstruction StreamsMISDMIMDMultiple
17SPMD Model (Single Program Multiple Data) Each processor executes the same program asynchronouslySynchronization takes place only when processors need to exchange dataSPMD is extension of SIMD (relax synchronized instruction execution)SPMD is restriction of MIMD (use only one source/object)
18Parallel Processing Terminology Embarassingly Parallel:applications which are trivial to parallelizelarge amounts of independent computationLittle communicationData Parallelism:model of parallel computing in which a single operation can be applied to all data elements simultaneouslyamenable to SIMD or SPMD style of computationControl Parallelism:many different operations may be executed concurrentlyrequire MIMD/SPMD style of computation
19Parallel Processing Terminology Scalability:If the size of problem is increased, number of processors that can be effectively used can be increased (i.e. there is no limit on parallelism).Cost of scalable algorithm grows slowly as input size and the number of processors are increased.Data parallel algorithms are more scalable than control parallel algorithmsGranularity:fine grain machines: employ massive number of weak processors each with small memorycoarse grain machines: smaller number of powerful processors each with large amounts of memory
20Shared Memory Machines Shared Address Spaceprocess(thread)Memory is globally shared, therefore processes (threads) see single addressspaceCoordination of accesses to locations done by use of locks provided bythread librariesExample Machines: Sequent, Alliant, SUN Ultra, Dual/Quad Board Pentium PCExample Thread Libraries: POSIX threads, Linux threads.
21Shared Memory Machines can be classified as:UMA: uniform memory accessNUMA: nonuniform memory accessbased on the amount of time a processor takes to access local and global memory.PM..PM..Inter-connectionnetworkInter-connectionnetworkM..P..Inter-connectionnetwork/or BUSM..(a)(b)(c)
22Distributed Memory Machines NetworkprocessMEach processor has its own local memory (not directly accessible by others)Processors communicate by passing messages to each otherExample Machines: IBM SP2, Intel Paragon, COWs (cluster of workstations)Example Message Passing Libraries: PVM, MPI
23Beowulf Clusters Use COTS, ordinary PCs and networking equipment Has the best price/performance ratioPC cluster
48Grid Computingprovide access to computing power and various resources just like accessing electrical power from electrical gridAllows coupling of geographically distributed resourcesProvide inexpensive access to resources irrespective of their physical location or access pointInternet & dedicated networks can be used to interconnect distributed computational resources and present them as a single unified resourceResources: supercomputers, clusters, storage systems, data resources, special devices
49Grid Computingthe GRID is, in effect, a set of software tools, which when combined with hardware, would let users tap processing power off the Internet as easily as the electrical power can be drawn from the electricty grid.Examples of Grid projects:- : search for extraterrestial intelligence- Entropia : company to broker processing power of idle computers, about 30,000 volunteer computers and total processing power 1 Tflop.: sifts astronomical data for pulsars- : protein folding- : population dynamics
50Seti@home Project Screen-saver program Sifts through signals recorded by the giant Arecibo radio telescope in Puerto Rico3 million people downloaded screen saver and run it.Program periodically prompts its host to retrieve a new chunk of data from the Internet and sends latest processed results back to SETI.Equivalent of more than 600,000 years of PC processing time has already clocked up.
51More Grid ProjectsGriPhyN: grid developed by consortium of American labs for physics projectsEarth System Grid: make huge climate simulations spanning hundreds of years.Earthquake Engineering Simulation Grid:Particle Physics Data Grid:Information Power Grid: supported by NASA for massive engineering calculationsDataGrid : European, coordinated by CERN. Aim is to develop middleware for research projects in biological sciences, earth observation and high energy physics.
52Gordon Bell & Jim Gray on “What’s next in High Performance Computing” Beowulf ’s economics and sociology are poised to kill off the other architectural linesComputational Grid can federate systems into supercomputers far beyond the power of any current computing centerThe centers will become super-data and super-application centersClusters (currently) perform poorly on applications that require large shared memory
53Gordon Bell & Jim Gray on “What’s next in High Performance Computing” Now individuals and laboratories can assemble and incrementally grow any-size super-computer anywhere in the world.By 2010, the cluster is likely to be the principal computing structure.does not run Linpack, so does not qualify in the top500 list. But avarages 13 Tflops making it more powerful than the top 3 of top500 machines combined.GRID and P2P computing using the Internet is likely to remain the world’s most powerful supercomputer.
54Gordon Bell & Jim Gray on “What’s next in High Performance Computing” Concerned that traditional supercomputer architecture is dead and a supercomputer mono-culture is being born.Recommend increased investment in peta-scale distributed databases.By 2010, the cluster is likely to be the principal computing structure.Research programs that stimulate cluster understanding and training are a good investment for laboratories that depend on highest performance machines.