CS591x -Cluster Computing and Parallel Programming Parallel Computer Architecture and Software Models
It all about performance Greater performance is the reason for parallel computing Many types of scientific and engineering programs are too large and too complex for traditional uniprocessors Such large problems are common is – Ocean modeling, weather modeling, astrophysics, solid state physics, power systems….
FLOPS – a measure of performance FLOPS – Floating Point Operations per Second … a measure of how much computation can be done in a certain amount of time MegaFLOPS – MFLOPS - 106 FLOPS GigaFLOPS – GFLOPS – 109 FLOPS TeraFLOPS – TFLOPS – 1012 FLOPS PetaFLOPS – PFLOPS – 1015 FLOPS
How fast … Cray 1 - ~150 MFLOPS Pentium 4 – 3-6 GFLOPS IBM’s BlueGene - +70 TFLOPS PSC’s Big Ben – 10 TFLOPS Humans --- it depends as calculators – 0.001 MFLOPS as information processors – 10PFLOPS
FLOPS vs. MIPS FLOPS only concerned with floating pointing calculations other performance issues memory latency cache performance I/O capacity …
See… www.Top500.org biannual performance reports and … rankings of the fastest computers in the world
Performance Speedup(n processors) = time(1 processor)/time(n processors) ** Culler, Singh and Gupta, Parallel Computing Architecture, A Hardware/Software Approach
Consider… from: www.lib.utexas.edu/maps/indian_ocean.html
… a model of the Indian Ocean - 73,000,000 square kilometer One data point per 100 meters 7,300,000,000 surface points Need to model the ocean at depth – say every 10 meters up to 200 meters 20 depth data points Every 10 minutes for 4 hours – 24 time steps
So – 73 x 106 (points on the surface) x 102 (points per sq. km) x 20 points per sq km of depth) x 24 (time steps) 3,504,000,000,000 data points in the model grid Suppose 100 instruction per grid point 350,400,000,000,000 instructions in model
Then - Imagine that you have a computer that can run 1 billion (109)instructions per second 3.504 x 1014 / 109 = 35040 seconds or 9.7 hours
But – On a 10 teraflops computer – 3.504 x 1014 / 1013 = 35.0 seconds
Gaining performance Pipelining More instructions –faster More instructions in execution at the same time in a single processor Not usually an attractive strategy these days – why?
Instruction Level Parallelism (ILP) based on the fact that many instructions do not depend on instructions that are before them… Processor has extra hardware to execute several instructions at the same time …multiple adders…
Pipelining and ILP not the solution to our problem – why? near incremental improvements in performance been done already we need orders of magnitude improvements in performance
Gaining Performance Vector Processors Scientific and Engineering computations are often vector and matrix operations graphic transformations – i.e. shift object x to the right Redundant arithmetic hardware and vector registers to operate on an entire vector in one step (SIMD)
Gaining Performance Vector Processors Declining popularity for a while – Hardware expensive Popularity returning – Applications – science, engineering, cryptography, media/graphics Earth Simulator
Parallel Computer Architecture Shared Memory Architectures Distributed Memory
Shared Memory Systems Multiple processors connected to/share the same pool of memory SMP Every processor has, potentially, access to and control of every memory location
Shared Memory Computers Processor Processor Processor Memory Processor Processor Processor
Shared Memory Computers Processor Processor Processor
Shared Memory Computer Switch Processor Processor Processor
Share Memory Computers SGI Origin2000 – at NCSA Balder 256 250mhz R10000 processors 128 Gbyte Memory
Shared Memory Computers Rachel at PSC 64 1.15 Ghz EV7 processors 256 Gbytes of shared memory
Distributed Memory Systems Multiple processors each with their own memory Interconnected to share/exchange data, processing Modern architectural approach to supercomputers Supercomputers and Clusters similar
Clusters – distributed memory Processor Processor Processor Interconnect Processor Processor Processor Memory Memory Memory
Cluster Distributed Memory with SMP Proc1 Proc2 Proc1 Proc2 Proc1 Proc2 Interconnect Proc1 Proc2 Proc1 Proc2 Proc1 Proc2 Memory Memory Memory
Distributed Memory Supercomputer BlueGene/L DOE/IBM 0.7 Ghz PowerPC 440 32768 Processors 70 Teraflops
Distributed Memory Supercomputer Thunder at LLNL Number 5 20 Teraflops 1.4 Ghz Itanium processors 4096 processors
Grid Computing Systems What is a Grid Means different things to different people Distributed Processors Around campus Around the state Around the world
Grid Computing Systems Widely distributed Loosely connected (i.e. Internet) No central management
Grid Computing Systems Connected Clusters/other dedicated scientific computers I2/Abilene
Grid Computer Systems Harvested Idle Cycles Internet Control/Scheduler
Grid Computing Systems Dedicated Grids TeraGrid Sabre NASA Information Power Grid Cycle Harvesting Grids Condor *GlobalGridForum (Parabon) Seti@home
Let’s revisit speedup… we can achieve speedup (theoretically) by using more processors,… but, of factors may limit speedup… Interprocessor communications Interprocess synchronization Load balance
Amdahl’s Law According to Amdahl’s Law… Speedup = 1/(S + (1-S)/N) where S is the purely sequential part of the program N is the number of processors
Amdahl’s Law What does it mean – Amdahl’s law says – Part of a program can is parallelizable Part of the program must be sequential (S) Amdahl’s law says – Speedup is constrained by the portion of the program that must remain sequential relative to the part that is parallelized. Note: If S is very small – “embarrassingly parallel problem”
Software models for parallel computing Shared Memory Distributed Memory Data Parallel
Flynn’s Taxonomy Single Instruction/Single Data - SISD Multiple Instruction/Single Data - MISD Single Instruction/Multiple Data - SIMD Multiple Instruction/Multiple Data - MIMD Single Program/Multiple Data - SPMD
Next Cluster Computer Architecture Linux