Computing & Information Science 4205 Effective Use of High Performance

Computing & Information Science 4205 Effective Use of High Performance Computing Steve Lantz [email protected] www.cac.cornell.edu/ slantz Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 1

Week 1 Lecture Notes Course Overview Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 2

Goals for CIS 4205 Introduction to HPC – Practical experience for your research Finding the parallelism in your work Measuring speedup & efficiency and the factors that affect it Writing & debugging parallel code (MPI & OpenMP) Exposure to using production HPC systems at Cornell Effective techniques for inherently (“embarrassingly”) parallel codes Critical analysis of current & future HPC solutions Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 3

A Little About Me Role at the Cornell Center for Advanced Computing – Senior Research Associate: consulting, training, advising, & participating – Involved in HPC around 25 years Background – Education – Experience Research interests – Numerical modeling and simulation – Fluid and plasma dynamics – Parallel computing Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 4

A Little About You* Fields of study and/or research interests Programming experience – C, C , Fortran, Others – Scripting languages Practical experiences – Ever written a program from scratch for your research? – Ever had to work with someone else’s code? – Which was harder? Why? HPC experience Your goals for this course * - (“A Little Bit Me, A Little Bit You” – Monkees, 1967) Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 5

Assignments Check the course website before every class – http://www.cac.cornell.edu/ slantz/CIS4205 Assignments are due on date specified Assignments should be emailed to me – [email protected] Assignments can be done on any HPC system – Windows, Linux, Macintosh OS X – HPC system must have MPI, OpenMP & batch scheduling system Access to CAC HPC systems is available Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 6

Connecting to CAC Resources All students’ Cornell NetIDs will be added the course account – Everyone will have the option to use CAC resources for assignments Accessing CAC machines – http://www.cac.cornell.edu/Documentation/Linux.asx – http://www.cac.cornell.edu/Documentation/Linux Poll: what’s your background? – Familiar with Linux? – Familiar with ssh and X Windows? – Comfortable with text editors (emacs, vi)? Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 7

Where and How to Find Me Physical and and virtual whereabouts – Office: 533 Frank H. T. Rhodes Hall – Phone: 4-8887 – Email: [email protected] Office hours by appointment Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 8

Introduction to High Performance Computing Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 9

Why HPC? Why Parallel Computing? Want to have best possible time-to-solution, minimize waiting Want to gain a competitive advantage As expected, processor clock speeds have flattened out – Not the end of Moore’s law: transistor densities are still doubling every 1.5 years – Clock speeds limited by power consumption, heat dissipation, current leakage – Bad news for mobile computers! Parallelism will be the path toward future performance gains – Trend is toward multi-core: put a cluster on a chip (in a laptop) – Goes well beyond microarchitectures that have multiple functional units Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 10

Evolution of HPC Centralized Big-Iron Mainframes – CRAYs – Vector Supercomputers Mini Computers Specialized Parallel Computers RISC MPPS NOWS Clusters Grids Clusters PCs RISC Workstations PCs Decentralized Collections 1970 Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 1980 1990 2000 11

The Cornell Programmer’s View Timeframe – hardware – parallel programming model Mid ’80’s – IBM mainframe with attached Floating Point Systems array processors – IBM and FPS compiler extensions Late ’80’s – Interconnected IBM 3090 mainframes featuring internal vector processors – Parallel VS FORTRAN and APF compilers Early ’90’s – IBM SP1, rack-mounted RS6000 workstations with POWER RISC processors networked via multistage crossbar switch; KSR-1 from Kendall Square Research – PVM, MPL, MPI message passing libraries; HPF/KAP directives; KSR compiler for ALLCACHE “virtual shared memory” Mid ’90’s – IBM SP2 featuring POWER2 and P2SC – MPI (not a compiler) Late ’90’s to present – Several generations of Dell HPC clusters (Velocity 1, 1 , 2, 3), quad or dual Intel Pentiums running Windows, Red Hat Linux – MPI plus OpenMP compiler directives for multithreading Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 12

Current HPC Platforms: COTS-Based Clusters COTS Commercial off-the-shelf Access Control File Server(s) Login Node(s) Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz Compute Nodes 13

Shared and Distributed Memory CPU1 core1 CPU2 core2 core1 core2 node1 node2 node3 FSB controller RAM Cluster interconnect Platform (node) Shared memory on each node Distributed memory across cluster Multi-core CPUs in clusters – two types of parallelism to consider Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 14

Shared and Distributed Memory CPU1 core1 CPU2 core2 core1 core2 node1 node2 node3 FSB controller RAM Cluster interconnect Platform (node) Shared memory on each node Distributed memory across cluster OpenMP Pthreads MPI MPI Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 15

OpenMP and MPI? In 2009? Strengths: Adherence to carefully defined specifications (not standards per se) Specifications still under active development Cross-platform, multi-OS, multi-language (e.g., pypar in Python) Wide acceptance Time-tested with large existing code base Useful for both data and task (functional) parallelism Weaknesses: Relatively low-level programming (though not as low as pthreads) Mindset taken from procedural languages (C, Fortran) Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 16

Where’s My Parallel Compiler? You’ve had it for years! “Serial” compilers produce code that takes advantage of parallelism at the top of the memory hierarchy Example: SSE(2/3/4) instructions operate on several floats or doubles simultaneously using special 128-bit-wide registers in Intel Xeons (vector processing) http://www.tomshardware.com/2006/06/26/xeon woodcrest preys on opteron/page9.html Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 17

Parallelism Inside the Intel Core In the Intel Core microarchitecture: 4 instructions per cycle Branch prediction Out-of-order execution 5 prefetchers 4MB L2 cache 3 128-bit SSE registers 1 SSE / cycle http://www.anandtech.com/cpuchipsets/showdoc.aspx?i 2748&p 4 Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 18

Why Not Crank Up the Clock? Because the CPU’s power consumption goes up like the cube of frequency! No wonder Intel tries so hard to boost the IPC Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 19

OK, What’s Next? Trend toward growing numbers of cores per processor die Moore’s Law still holds; transistor densities are still increasing Higher densities don’t translate into faster speeds due to: – Problems with heat dissipation – Hefty power requirements – Leakage current The “free lunch” of everincreasing clock speeds is over! Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz http://www.gotw.ca/publications/concurrency-ddj.htm 20

Signs of the Times IBM promotes BlueGene HPC line with 1000’s of low-frequency, lowpower chips (700 MHz PowerPCs) On 2/11/07, Intel announces successful tests of an 80-core research processor – “teraflops on a chip” Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 21

Implications for Programmers Concurrency will no longer be just desirable, it will be essential – Previously it was the economical path to greater performance – Now it has become the physically reasonable path Compilers (still) aren't the answer – Degree of concurrency depends on algorithm choices – Need for high-level creation (as opposed to mere identification) of concurrent code sections Improved programming languages could make life easier, but nothing has caught on yet Some newer languages (Java, C ) do have mechanisms for concurrency built in but kind of clumsy to use In the final analysis: TANSTAAFL Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 22

Conclusions Future processor technology will drive programmers to create and exploit concurrency in their software to get performance Some problems are inherently not parallelizable; what then? – – – – “9 women can't produce a baby in one month” But what if the goal isn't just to produce one baby, but many? “Embarrassing parallelism” isn’t so embarrassing any more Examples: optimization of a design; high-level Monte Carlo simulation Coding for efficiency and performance optimization will get more, not less, important – Not all performance gains need to come from high-level parallelism – Nevertheless, parallelism needs to be designed into codes, preferably from the beginning Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 23

Parallel Computing: Types of Parallelism Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 24

Parallel Computing: Definitions As we have seen, HPC necessarily relies on parallel computing Parallel computing – Involves the use of multiple processors simultaneously to reduce the time needed to solve a single computational problem. Examples of fields where this is important: – – – – Climate modeling, weather forecasting Aircraft and ship design Cosmology, simulations of the evolution of stars and galaxies Molecular dynamics and electronic (quantum) structure Parallel programming – Is writing code in a language (plus extensions) that allows you to explicitly indicate how different portions of the computation may be executed concurrently Therefore, the first step in parallel programming is to identify the parallelism in the way your problem is being solved (algorithm) Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 25

Data Parallelism Definition: when independent tasks can apply the same operation to different elements of the data set at the same time. Examples: 2 brothers mow the lawn 8 farmers paint a barn A B B B C Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 26

Data Parallelism Partitions of the Data Can Be Processed Simultaneously // initialize array values to 1 a[] 1; b[] 1; c[] 1; for (i 0; i 3; i ) { a[i] b[i] c[i]; } // Serial Execution i 0 (a[0] 2) a[0] b[0] c[0]; i 1 (a[1] 2) a[1] b[1] c[1]; i 2 (a[2] 2) a[2] b[2] c[2]; // Parallel Execution i 0 (a[0] 2) i 1 (a[1] 2) i 2 (a[2] 2) a[0] b[0] c[0]; a[1] b[1] c[1]; a[2] b[2] c[2]; Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 27

Data Parallelism MPI Example #include stdio.h #include mpi.h #include malloc.h void main(int argc, char **argv ) { int myid, numprocs; int i; int *a,*b,*c; MPI Status status; MPI Init(&argc, &argv); MPI Comm size(MPI COMM WORLD,&numprocs); MPI Comm rank(MPI COMM WORLD,&myid); a (int *) malloc(numprocs*sizeof(int)); b (int *) malloc(numprocs*sizeof(int)); c (int *) malloc(numprocs*sizeof(int)); for (i 0;i numprocs;i ) { // initialize array values to i a[i] i; b[i] i; c[i] i; } a[myid] b[myid] c[myid]; printf("a[%d] %d\n",myid,a[myid]); MPI Finalize(); } Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz mpiexec –n 8 dp1.exe 3: a[3] 6 4: a[4] 8 5: a[5] 10 7: a[7] 14 0: a[0] 0 2: a[2] 4 6: a[6] 12 1: a[1] 2 28

Functional Parallelism Definition: when independent tasks can apply different operations to the same (or different) data elements at the same time. Examples: 2 brothers do yard work (1 rakes, 1 mows) 8 farmers build a barn B A C D E Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 29

Functional Parallelism Partitions of the Program Can Execute Simultaneously // initialize values 0-9 a[] i; b[] i; // These different operations can happen at the same time for (i 0; i 10; i ) { c[i] a[i] b[i]; } for (i 0; i 10; i ) { d[i] a[i] * b[i]; } // This part requires solutions from above for (i 0; i 10; i ) { e[i] d[i] - c[i]; } Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 30

Functional Parallelism MPI Example #include stdio.h #include mpi.h #include malloc.h void main(int argc, char **argv ) { int myid, numprocs; int i; int iter 10; int *a,*b,*c,*d,*e; MPI Status status; MPI Init(&argc, &argv); MPI Comm size(MPI COMM WORLD,&numprocs); MPI Comm rank(MPI COMM WORLD,&myid); if (numprocs 3) { printf("ERROR: This example requires 3 processes\n"); } else { a (int *) malloc(iter*sizeof(int)); b (int *) malloc(iter*sizeof(int)); c (int *) malloc(iter*sizeof(int)); d (int *) malloc(iter*sizeof(int)); for (i 0; i iter; i ) { a[i] i; b[i] i; } Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz if (myid 0) { MPI Recv(c,10,MPI INT,1,0,MPI COMM WORLD,&status); MPI Recv(d,10,MPI INT,2,0,MPI COMM WORLD,&status); e (int *) malloc(iter*sizeof(int)); for (i 0; i iter; i ) { e[i] d[i] - c[i]; printf("e[%d] %d\n",i,e[i]); } } else if (myid 1) { for (i 0; i iter; i ) { c[i] a[i] b[i]; } MPI Send(c,10,MPI INT,0,0,MPI COMM WORLD); } else if (myid 2) { for (i 0; i iter; i ) { d[i] a[i] * b[i]; } MPI Send(d,10,MPI INT,0,0,MPI COMM WORLD); } else { printf("Process id %d not needed\n",myid); } } MPI Finalize(); } 31

Task Parallelism Definition: when independent “Worker” tasks can perform functions that do not need to communicate with each other, only with a “Master” or “Manager” process. Such tasks are often called “Embarrassingly Parallel” because they can be parallelized with little extra work or thought. Examples: Independent Monte Carlo Simulations ATM Transactions A B Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz C D 32

Task Parallelism Independent Tasks are Distributed as Workers are Available // initialize values 0-99 a[] i; b[] i; // Send each idle worker an index of a[] & b[] to add and return the sum // and continue while there is still work to be done while (i 100) { // find an idle worker & send it value of i (index) to add // receive back summed values } Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 33

Pipeline Parallelism Definition: each task works on one stage in a sequence of stages. The output of one stage is the input of the next. (Note: This works best when each stage takes the same amount of time to complete) Examples: Assembly lines Computing partial sums T0 T 1 A B C Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz i T2 T3 T4 T5 T6 T7 T8 i 1 i 2 i 3 i 4 i 5 i 6 i i 1 i 2 i 3 i 4 i 5 i 6 i i 1 i 2 i 3 i 4 i 5 i 6 i i 1 i 2 i 3 i 4 i 5 T9 i 6 34

Pipeline Parallelism // Process 0 initializes a[] & b[] for (i 0; i 3; i ) { a[i] i; b[i] 0; } b[0] a[0]; // Process 0 sends a & b to Process 1 // Process 1 receives a & b from Process 0 b[1] b[0] a[1]; // Process 1 sends a & b to Process 2 // Process 2 receives a & b from Process 1 b[2] b[1] a[2]; // Process 2 sends a & b to Process 3 // Process 3 receives a & b from Process 2 b[3] b[2] a[3]; Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 35

Tightly vs. Loosely Coupled Parallelism Tightly Coupled Parallel Approaches Parallel tasks must exchange data during the computation Loosely Coupled Parallel Approaches Parallel tasks can complete independent of each other for (int i 0; i n; i ) { for (int j 0; j m; j ) { //Perform Calculation Here } // for j } // for i Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 36

Tightly Coupled Example: Strong Data Dependencies for (int i 0; i n; i ) i i 1 i 2 i 3 i n ;j 0 tj in r( fo ;j 0 tj in r( fo ;j 0 tj in r( fo ;j 0 tj in r( fo ;j 0 tj in r( fo ) ;j m ) ;j m ) ;j m ) ;j m ) ;j m Workers Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 37

Loosely Coupled Example: Master-Worker Codes i 1 i 2 i Master for (int i 0; i n; i ) i 3 i 4 i n for (int j 0; j m; j ) for (int j 0; j m; j ) for (int j 0; j m; j ) for (int j 0; j m; j ) for (int j 0; j m; j ) Workers Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 38

HPC Examples at Cornell: Parallel Computing and Data Intensive Computing Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 39

Parallel Computing Examples Computational Biology Service Unit http://cbsu.tc.cornell.edu/index.htm Cornell Fracture Group http://www.cfg.cornell.edu Computational Finance http://www.orie.cornell.edu/orie/manhattan/ Cornell Institute for Social and Economic Research http://www.ciser.cornell.edu/ Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 40

Data Intensive Computing Applications Modern Research is Producing Massive Amounts of Data – – – – – – – Microscopes Telescopes Gene Sequencers Mass Spectrometers Satellite & Radar Images Distributed Weather Sensors High Performance Computing (especially HPC Clusters) Research Communities Rely on Distributed Data Sources – – – Collaboration Virtual Laboratories Laboratory Information Management Systems (LIMS) New Management and Usage Issues – – – – Security Reliability/Availability Manageability Data Locality – You can’t ftp a petabyte to your laptop . Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 41

Data Intensive Computing Examples Arecibo - World’s Largest Radiotelescope Johannes Gehrke, Jim Cordes, David Lifka Serving Astronomy Data via SQL Server and Web Services http://arecibo.tc.cornell.edu/PALFA http://www.cs.cornell.edu/johannes/ Cornell Fracture Group Tony Ingraffea Serving Finite Element Models via SQL Server & Web Services http://www.cfg.cornell.edu/ Physically Accurate Imagery Steve Marschner http://www.cs.cornell.edu/ srm/ The Structure and Evolution of the Web William Arms http://www.cs.cornell.edu/wya/ Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 42

High Performance Computing Architectures Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 43

Flynn’s Taxonomy Classification Scheme for Parallel Computers Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz Single Multiple Single SISD SIMD Multiple Instruction Stream Data Stream MISD MIMD 44

Examples from Flynn’s Categories SISD – Single Instruction Stream, Single Data Stream – Common uniprocessor machines SIMD – Single Instruction Stream, Multiple Data Streams – Processor arrays (including GPUs) & pipelined vector processors MISD – Multiple Instruction Streams, Single Data Stream – Systolic arrays: think data pump or pumping-heart model (not many built) MIMD – Multiple Instruction Streams, Multiple Data Streams – Multiprocessors and multicomputers Multiprocessor: multi-CPU computer with shared memory – SMP: Symmetric MultiProcessor (uniform memory access) – NUMA: Non Uniform Memory Access multiprocessor Multicomputer: team of computers with distributed CPUs and memory – Must have external interconnect between “nodes” Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 45

Shared vs. Switched Media bus switch processors processors SMP: not scalable, bus becomes congested as processors are added NUMA, Multicomputer: switch can grow with number of processors Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 46

2D Mesh and Torus Topologies Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 47

Hypercube Topology Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 48

Tree, Fat Tree, and Hypertree Topologies Fat tree if links get wider toward the top Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 49

Clos Network: Equal to a Full Crossbar Switch, Better Than a Hypertree (Fewer Hops) Generally n m, so inputs and outputs can be bundled into the same cable and plug into a single switch port Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 50

Comparison of Switched Media Type Latency Bandwidth Cost Gigabit Ethernet 1 msec 0.1 gigabyte/sec 10 Gigabit Ethernet 100 sec 1.0 gigabyte/sec QDR InfiniBand 1 sec 3.6 gigabyte/sec Mellanox 36-port InfiniBand switch Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 51

Computing Concepts Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 52

Single Processor Memory Hierarchy Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 53

Creating and Running Software Compiler – Produces object code: from myprog.c, creates myprog.o (Windows: .obj) Linker – Produces complete executable, including object code imported from libraries Shared Objects (.so) and Dynamic Load Libraries (Windows DLLs) – These are loaded at runtime: the link step inserts instructions on what to import – If a shared object is loaded, a single copy can be used by multiple processes Process – A running executable: the OS controls multitasking of processes via scheduling Virtual Memory – “Address space” available to a running process, addresses can be 32- or 64-bit Paging (to Disk) – Physical RAM has been exceeded: requested data are not in any cache (cache miss) or in RAM (page fault) must be loaded from swap space on a hard drive Steve Lantz Computing and Information Science 4205 www.cac.cornell.edu/ slantz 54