GPU Basics GPU Basics S
Total Page:16
File Type:pdf, Size:1020Kb

Load more
Recommended publications
-
Cimple: Instruction and Memory Level Parallelism a DSL for Uncovering ILP and MLP
Cimple: Instruction and Memory Level Parallelism A DSL for Uncovering ILP and MLP Vladimir Kiriansky, Haoran Xu, Martin Rinard, Saman Amarasinghe MIT CSAIL {vlk,haoranxu510,rinard,saman}@csail.mit.edu Abstract Processors have grown their capacity to exploit instruction Modern out-of-order processors have increased capacity to level parallelism (ILP) with wide scalar and vector pipelines, exploit instruction level parallelism (ILP) and memory level e.g., cores have 4-way superscalar pipelines, and vector units parallelism (MLP), e.g., by using wide superscalar pipelines can execute 32 arithmetic operations per cycle. Memory and vector execution units, as well as deep buffers for in- level parallelism (MLP) is also pervasive with deep buffering flight memory requests. These resources, however, often ex- between caches and DRAM that allows 10+ in-flight memory hibit poor utilization rates on workloads with large working requests per core. Yet, modern CPUs still struggle to extract sets, e.g., in-memory databases, key-value stores, and graph matching ILP and MLP from the program stream. analytics, as compilers and hardware struggle to expose ILP Critical infrastructure applications, e.g., in-memory databases, and MLP from the instruction stream automatically. key-value stores, and graph analytics, characterized by large In this paper, we introduce the IMLP (Instruction and working sets with multi-level address indirection and pointer Memory Level Parallelism) task programming model. IMLP traversals push hardware to its limits: large multi-level caches tasks execute as coroutines that yield execution at annotated and branch predictors fail to keep processor stalls low. The long-latency operations, e.g., memory accesses, divisions, out-of-order windows of hundreds of instructions are also or unpredictable branches. -
Introduction to Multi-Threading and Vectorization Matti Kortelainen Larsoft Workshop 2019 25 June 2019 Outline
Introduction to multi-threading and vectorization Matti Kortelainen LArSoft Workshop 2019 25 June 2019 Outline Broad introductory overview: • Why multithread? • What is a thread? • Some threading models – std::thread – OpenMP (fork-join) – Intel Threading Building Blocks (TBB) (tasks) • Race condition, critical region, mutual exclusion, deadlock • Vectorization (SIMD) 2 6/25/19 Matti Kortelainen | Introduction to multi-threading and vectorization Motivations for multithreading Image courtesy of K. Rupp 3 6/25/19 Matti Kortelainen | Introduction to multi-threading and vectorization Motivations for multithreading • One process on a node: speedups from parallelizing parts of the programs – Any problem can get speedup if the threads can cooperate on • same core (sharing L1 cache) • L2 cache (may be shared among small number of cores) • Fully loaded node: save memory and other resources – Threads can share objects -> N threads can use significantly less memory than N processes • If smallest chunk of data is so big that only one fits in memory at a time, is there any other option? 4 6/25/19 Matti Kortelainen | Introduction to multi-threading and vectorization What is a (software) thread? (in POSIX/Linux) • “Smallest sequence of programmed instructions that can be managed independently by a scheduler” [Wikipedia] • A thread has its own – Program counter – Registers – Stack – Thread-local memory (better to avoid in general) • Threads of a process share everything else, e.g. – Program code, constants – Heap memory – Network connections – File handles -
Scheduling Task Parallelism on Multi-Socket Multicore Systems
Scheduling Task Parallelism" on Multi-Socket Multicore Systems" Stephen Olivier, UNC Chapel Hill Allan Porterfield, RENCI Kyle Wheeler, Sandia National Labs Jan Prins, UNC Chapel Hill The University of North Carolina at Chapel Hill Outline" Introduction and Motivation Scheduling Strategies Evaluation Closing Remarks The University of North Carolina at Chapel Hill ! Outline" Introduction and Motivation Scheduling Strategies Evaluation Closing Remarks The University of North Carolina at Chapel Hill ! Task Parallel Programming in a Nutshell! • A task consists of executable code and associated data context, with some bookkeeping metadata for scheduling and synchronization. • Tasks are significantly more lightweight than threads. • Dynamically generated and terminated at run time • Scheduled onto threads for execution • Used in Cilk, TBB, X10, Chapel, and other languages • Our work is on the recent tasking constructs in OpenMP 3.0. The University of North Carolina at Chapel Hill ! 4 Simple Task Parallel OpenMP Program: Fibonacci! int fib(int n)! {! fib(10)! int x, y;! if (n < 2) return n;! #pragma omp task! fib(9)! fib(8)! x = fib(n - 1);! #pragma omp task! y = fib(n - 2);! #pragma omp taskwait! fib(8)! fib(7)! return x + y;! }! The University of North Carolina at Chapel Hill ! 5 Useful Applications! • Recursive algorithms cilksort cilksort cilksort cilksort cilksort • E.g. Mergesort • List and tree traversal cilkmerge cilkmerge cilkmerge cilkmerge cilkmerge cilkmerge • Irregular computations cilkmerge • E.g., Adaptive Fast Multipole cilkmerge cilkmerge -
Concurrent Cilk: Lazy Promotion from Tasks to Threads in C/C++
Concurrent Cilk: Lazy Promotion from Tasks to Threads in C/C++ Christopher S. Zakian, Timothy A. K. Zakian Abhishek Kulkarni, Buddhika Chamith, and Ryan R. Newton Indiana University - Bloomington, fczakian, tzakian, adkulkar, budkahaw, [email protected] Abstract. Library and language support for scheduling non-blocking tasks has greatly improved, as have lightweight (user) threading packages. How- ever, there is a significant gap between the two developments. In previous work|and in today's software packages|lightweight thread creation incurs much larger overheads than tasking libraries, even on tasks that end up never blocking. This limitation can be removed. To that end, we describe an extension to the Intel Cilk Plus runtime system, Concurrent Cilk, where tasks are lazily promoted to threads. Concurrent Cilk removes the overhead of thread creation on threads which end up calling no blocking operations, and is the first system to do so for C/C++ with legacy support (standard calling conventions and stack representations). We demonstrate that Concurrent Cilk adds negligible overhead to existing Cilk programs, while its promoted threads remain more efficient than OS threads in terms of context-switch overhead and blocking communication. Further, it enables development of blocking data structures that create non-fork-join dependence graphs|which can expose more parallelism, and better supports data-driven computations waiting on results from remote devices. 1 Introduction Both task-parallelism [1, 11, 13, 15] and lightweight threading [20] libraries have become popular for different kinds of applications. The key difference between a task and a thread is that threads may block|for example when performing IO|and then resume again. -
Parallel Computing a Key to Performance
Parallel Computing A Key to Performance Dheeraj Bhardwaj Department of Computer Science & Engineering Indian Institute of Technology, Delhi –110 016 India http://www.cse.iitd.ac.in/~dheerajb Dheeraj Bhardwaj <[email protected]> August, 2002 1 Introduction • Traditional Science • Observation • Theory • Experiment -- Most expensive • Experiment can be replaced with Computers Simulation - Third Pillar of Science Dheeraj Bhardwaj <[email protected]> August, 2002 2 1 Introduction • If your Applications need more computing power than a sequential computer can provide ! ! ! ❃ Desire and prospect for greater performance • You might suggest to improve the operating speed of processors and other components. • We do not disagree with your suggestion BUT how long you can go ? Can you go beyond the speed of light, thermodynamic laws and high financial costs ? Dheeraj Bhardwaj <[email protected]> August, 2002 3 Performance Three ways to improve the performance • Work harder - Using faster hardware • Work smarter - - doing things more efficiently (algorithms and computational techniques) • Get help - Using multiple computers to solve a particular task. Dheeraj Bhardwaj <[email protected]> August, 2002 4 2 Parallel Computer Definition : A parallel computer is a “Collection of processing elements that communicate and co-operate to solve large problems fast”. Driving Forces and Enabling Factors Desire and prospect for greater performance Users have even bigger problems and designers have even more gates Dheeraj Bhardwaj <[email protected]> -
Task Parallelism Bit-Level Parallelism
Parallel languages as extensions of sequential ones Alexey A. Romanenko [email protected] What this section about? ● Computers. History. Trends. ● What is parallel program? ● What is parallel programming for? ● Features of parallel programs. ● Development environment. ● etc. Agenda 1. Sequential program 2. Applications, required computational power. 3. What does parallel programming for? 4. Parallelism inside ordinary PC. 5. Architecture of modern CPUs. 6. What is parallel program? 7. Types of parallelism. Agenda 8. Types of computational installations. 9. Specificity of parallel programs. 10.Amdahl's law 11.Development environment 12.Approaches to development of parallel programs. Cost of development. 13.Self-test questions History George Boole Claude Elwood Shannon Alan Turing Charles Babbage John von Neumann Norbert Wiener Henry Edward Roberts Sciences ● Computer science is the study of the theoretical foundations of information and computation, and of practical techniques for their implementation and application in computer systems. ● Cybernetics is the interdisciplinary study of the structure of regulatory system Difference machine Arithmometer Altair 8800 Computer with 8-inch floppy disk system Sequential program A program perform calculation of a function F = G(X) for example: a*x2+b*x+c=0, a != 0. x1=(-b-sqrt(b2-4ac))/(2a), x2=(-b+sqrt(b2-4ac))/(2a) Turing machine Plasma modeling N ~ 106 dX ~ F dT2 j j F ~ sum(q, q ) j i i j Complexity ~ O(N*N) more then 1012 * 100...1000 operations Resource consumable calculations ● Nuclear/Gas/Hydrodynamic -
Control Replication: Compiling Implicit Parallelism to Efficient SPMD with Logical Regions
Control Replication: Compiling Implicit Parallelism to Efficient SPMD with Logical Regions Elliott Slaughter Wonchan Lee Sean Treichler Stanford University Stanford University Stanford University SLAC National Accelerator Laboratory [email protected] NVIDIA [email protected] [email protected] Wen Zhang Michael Bauer Galen Shipman Stanford University NVIDIA Los Alamos National Laboratory [email protected] [email protected] [email protected] Patrick McCormick Alex Aiken Los Alamos National Laboratory Stanford University [email protected] [email protected] ABSTRACT 1 for t = 0, T do We present control replication, a technique for generating high- 1 for i = 0, N do −− Parallel 2 for i = 0, N do −− Parallel performance and scalable SPMD code from implicitly parallel pro- 2 for t = 0, T do 3 B[i] = F(A[i]) grams. In contrast to traditional parallel programming models that 3 B[i] = F(A[i]) 4 end 4 −− Synchronization needed require the programmer to explicitly manage threads and the com- 5 for j = 0, N do −− Parallel 5 A[i] = G(B[h(i)]) munication and synchronization between them, implicitly parallel 6 A[j] = G(B[h(j)]) 6 end programs have sequential execution semantics and by their nature 7 end 7 end avoid the pitfalls of explicitly parallel programming. However, with- 8 end (b) Transposed program. out optimizations to distribute control overhead, scalability is often (a) Original program. poor. Performance on distributed-memory machines is especially sen- F(A[0]) . G(B[h(0)]) sitive to communication and synchronization in the program, and F(A[1]) . G(B[h(1)]) thus optimizations for these machines require an intimate un- . -
Regent: a High-Productivity Programming Language for Implicit Parallelism with Logical Regions
REGENT: A HIGH-PRODUCTIVITY PROGRAMMING LANGUAGE FOR IMPLICIT PARALLELISM WITH LOGICAL REGIONS A DISSERTATION SUBMITTED TO THE DEPARTMENT OF COMPUTER SCIENCE AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY Elliott Slaughter August 2017 © 2017 by Elliott David Slaughter. All Rights Reserved. Re-distributed by Stanford University under license with the author. This work is licensed under a Creative Commons Attribution- Noncommercial 3.0 United States License. http://creativecommons.org/licenses/by-nc/3.0/us/ This dissertation is online at: http://purl.stanford.edu/mw768zz0480 ii I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy. Alex Aiken, Primary Adviser I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy. Philip Levis I certify that I have read this dissertation and that, in my opinion, it is fully adequate in scope and quality as a dissertation for the degree of Doctor of Philosophy. Oyekunle Olukotun Approved for the Stanford University Committee on Graduate Studies. Patricia J. Gumport, Vice Provost for Graduate Education This signature page was generated electronically upon submission of this dissertation in electronic format. An original signed hard copy of the signature page is on file in University Archives. iii Abstract Modern supercomputers are dominated by distributed-memory machines. State of the art high-performance scientific applications targeting these machines are typically written in low-level, explicitly parallel programming models that enable maximal performance but expose the user to programming hazards such as data races and deadlocks. -
A CPU/GPU Task-Parallel Runtime with Explicit Epoch Synchronization
TREES: A CPU/GPU Task-Parallel Runtime with Explicit Epoch Synchronization Blake A. Hechtman, Andrew D. Hilton, and Daniel J. Sorin Department of Electrical and Computer Engineering Duke University Abstract —We have developed a task-parallel runtime targeting CPUs are a poor fit for GPUs. To understand system, called TREES, that is designed for high why this mismatch exists, we must first understand the performance on CPU/GPU platforms. On platforms performance of an idealized task-parallel application with multiple CPUs, Cilk’s “work-first” principle (with no runtime) and then how the runtime’s overhead underlies how task-parallel applications can achieve affects it. The performance of a task-parallel application performance, but work-first is a poor fit for GPUs. We is a function of two characteristics: its total amount of build upon work-first to create the “work-together” work to be performed (T1, the time to execute on 1 principle that addresses the specific strengths and processor) and its critical path (T∞, the time to execute weaknesses of GPUs. The work-together principle on an infinite number of processors). Prior work has extends work-first by stating that (a) the overhead on shown that the runtime of a system with P processors, the critical path should be paid by the entire system at TP, is bounded by = ( ) + ( ) due to the once and (b) work overheads should be paid co- greedy o ff line scheduler bound [3][10]. operatively. We have implemented the TREES runtime A task-parallel runtime introduces overheads and, for in OpenCL, and we experimentally evaluate TREES purposes of performance analysis, we distinguish applications on a CPU/GPU platform. -
An Overview of Parallel Ccomputing
An Overview of Parallel Ccomputing Marc Moreno Maza University of Western Ontario, London, Ontario (Canada) CS2101 Plan 1 Hardware 2 Types of Parallelism 3 Concurrency Platforms: Three Examples Cilk CUDA MPI Hardware Plan 1 Hardware 2 Types of Parallelism 3 Concurrency Platforms: Three Examples Cilk CUDA MPI Hardware von Neumann Architecture In 1945, the Hungarian mathematician John von Neumann proposed the above organization for hardware computers. The Control Unit fetches instructions/data from memory, decodes the instructions and then sequentially coordinates operations to accomplish the programmed task. The Arithmetic Unit performs basic arithmetic operation, while Input/Output is the interface to the human operator. Hardware von Neumann Architecture The Pentium Family. Hardware Parallel computer hardware Most computers today (including tablets, smartphones, etc.) are equipped with several processing units (control+arithmetic units). Various characteristics determine the types of computations: shared memory vs distributed memory, single-core processors vs multicore processors, data-centric parallelism vs task-centric parallelism. Historically, shared memory machines have been classified as UMA and NUMA, based upon memory access times. Hardware Uniform memory access (UMA) Identical processors, equal access and access times to memory. In the presence of cache memories, cache coherency is accomplished at the hardware level: if one processor updates a location in shared memory, then all the other processors know about the update. UMA architectures were first represented by Symmetric Multiprocessor (SMP) machines. Multicore processors follow the same architecture and, in addition, integrate the cores onto a single circuit die. Hardware Non-uniform memory access (NUMA) Often made by physically linking two or more SMPs (or multicore processors). -
Task Level Parallelism
Task Level Parallelism The topic of this chapter is thread-level parallelism. While, thread-level parallelism falls within the textbook’s classification of ILP and data parallelism. It also falls into a broader topic of parallel and distributed computing. In the next set of slides, I will attempt to place you in the context of this broader computation space that is called task level parallelism. Of course a proper treatment of parallel computing or distributed computing is worthy of an entire semester (or two) course of study. I can only give you a brief exposure to this topic. The text highlighted in green in these slides contain external hyperlinks. 1 / 14 Classification of Parallelism Software Sequential Concurrent Serial Some problem written as a se- Some problem written as a quential program (the MATLAB concurrent program (the O/S example from the textbook). example from the textbook). Execution on a serial platform. Execution on a serial platform. Parallel Some problem written as a se- Some problem written as a quential program (the MATLAB concurrent program (the O/S Hardware example from the textbook). example from the textbook). Execution on a parallel plat- Execution on a parallel plat- form. form. 2 / 14 Flynn’s Classification of Parallelism CU: control unit SM: shared memory DS1 PU MM PU: processor unit IS: instruction stream 1 1 MM: memory unit DS: data stream DS2 PU MM IS 2 2 CU IS SM IS DS CU PU MM DSn PUn MMm (a) SISD computer IS (b) SIMD computer IS1 IS1 IS1 IS1 IS1 DS1 CU PU DS CU PU MM 1 1 1 1 1 SM IS2 IS2 IS2 IS2 DS2 CU PU CU PU MM IS2 2 2 2 2 2 MM MM MM 1 2 m SM ISn ISn ISn ISn ISn DSn IS CUnPU n DS 2 CUnPU n MMm IS1 ISn (c) MISD computer (d) MIMD computer 3 / 14 Task Level Parallelism I Task Level Parallelism: organizing a program or computing solution into a set of processes/tasks/threads for simultaneous execution. -
Unified Parallel C for GPU Clusters: Language Extensions and Compiler Implementation
Unified Parallel C for GPU Clusters: Language Extensions and Compiler Implementation Li Chen1, Lei Liu1, Shenglin Tang1, Lei Huang2, Zheng Jing1, Shixiong Xu1, Dingfei Zhang1, Baojiang Shou1 1 Key Laboratory of Computer System and Architecture, Institute of Computing Technology, Chinese Academy of Sciences, China {lchen,liulei2007,tangshenglin,jingzheng,xushixiong, zhangdingfei, shoubaojiang}@ict.ac.cn; 2 Department of Computer Science, University of Houston; Houston, TX, USA [email protected] 5 Abstract. Unified Parallel C (UPC), a parallel extension to ANSI C, is designed for high performance computing on large-scale parallel machines. With General-purpose graphics processing units (GPUs) becoming an increasingly important high performance computing platform, we propose new language extensions to UPC to take advantage of GPU clusters. We extend UPC with hierarchical data distribution, revise the execution model of UPC to mix SPMD with fork-join execution model, and modify the semantics of upc_forall to reflect the data-thread affinity on a thread hierarchy. We implement the compiling system, including affinity-aware loop tiling, GPU code generation, and several memory optimizations targeting NVIDIA CUDA. We also put forward unified data management for each UPC thread to optimize data transfer and memory layout for separate memory modules of CPUs and GPUs. The experimental results show that the UPC extension has better programmability than the mixed MPI/CUDA approach. We also demonstrate that the integrated compile-time and runtime optimization is effective to achieve good performance on GPU clusters. 1 Introduction Following closely behind the industry-wide move from uniprocessor to multi-core and many-core systems, HPC computing platforms are undergoing another major change: from homogeneous to heterogeneous platform.