DOCSLIB.ORG

Parallel Programming: Performance

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Parallel Programming: Performance Introduction Parallel Programming: Rich space of techniques and issues • Trade off and interact with one another Performance Issues can be addressed/helped by software or hardware • Algorithmic or programming techniques • Architectural techniques Todd C. Mowry Focus here on performance issues and software techniques CS 418 • Point out some architectural implications January 20, 25 & 26, 2011 • Architectural techniques covered in rest of class –2– CS 418 Programming as Successive Refinement Outline Not all issues dealt with up front 1. Partitioning for performance Partitioning often independent of architecture, and done first 2. Relationshipp,y of communication, data locality and • View machine as a collection of communicating processors architecture – balancing the workload – reducing the amount of inherent communication 3. Orchestration for performance – reducing extra work • Tug-o-war even among these three issues For each issue: Then interactions with architecture • Techniques to address it, and tradeoffs with previous issues • Illustration using case studies • View machine as extended memory hierarchy • Application to g rid solver –extra communication due to archlhitectural interactions • Some architectural implications – cost of communication depends on how it is structured • May inspire changes in partitioning Discussion of issues is one at a time, but identifies tradeoffs 4. Components of execution time as seen by processor • Use examples, and measurements on SGI Origin2000 • What workload looks like to architecture, and relate to software issues –3– CS 418 –4– CS 418 Page 1 Partitioning for Performance Load Balance and Synch Wait Time Sequential Work 1. Balancing the workload and reducing wait time at synch Limit on speedup: Speedupproblem(p) < points Max Work on any Processor • Work includes data access and other costs 2. Reducing inherent communication • Not just equal work, but must be busy at same time 3. Reducing extra work Four parts to load balance and reducing synch wait time: Even these algorithmic issues trade off: • Minimize comm. => run on 1 processor => extreme load imbalance 1. Identify enough concurrency • Maximize load balance => random assignment of tiny tasks => no 2. Decide how to manage it control over communication • Good partition may imply extra work to compute or manage it 3. Determine the granularity at which to exploit it Goal is to compromise 4. Reduce serialization and cost of synchronization • Fortunately, often not difficult in practice –5– CS 418 –6– CS 418 Identifying Concurrency Identifying Concurrency (contd.) Techniques seen for equation solver: Function parallelism: • Loop structure, fundamental dependences, new algorithms • entire large tasks (procedures) that can be done in parallel Data Parallelism versus Function Parallelism • on same or different data Often see orthogonal levels of parallelism; e.g. VLSI routing • e.g. different independent grid computations in Ocean • pipelining, as in video encoding/decoding, or polygon rendering W1 W2 W3 • degree usually modest and does not grow with input size (a) • difficult to load balance Wire W2 expands to segments • often used to reduce synch between data parallel phases S21 S22 S23 S24 S25 S26 (b) Most scalable programs data parallel (per this loose definition) Segment S23 expands to routes • function parallelism reduces synch between data parallel phases (c) –7– CS 418 –8– CS 418 Page 2 Load Balance and Synch Wait Time Deciding How to Manage Concurrency Sequential Work Limit on speedup: Speedupproblem(p) < Static versus Dynamic techniques Max Work on any Processor Static: • Work includes data access and other costs • Algorithmic assignment based on input; won’t change • Not just equal work, but must be busy at same time • Low runtime overhead • Computation must be predictable Four parts to load balance and reducing synch wait time: • Preferable when applicable (except in multiprogrammed or heterogeneous environment) 1. Identify enough concurrency Dynamic: 2. Decide how to manage it • Adappunmnt at runtime to balance load 3. Determine the granularity at which to exploit it • Can increase communication and reduce locality • Can increase task management overheads 4. Reduce serialization and cost of synchronization –9– CS 418 –10– CS 418 Dynamic Assignment Dynamic Tasking with Task Queues Profile-based (semi-static): Centralized versus distributed queues • Profile work distribution at runtime, and repartition dynamically Task stealinggq with distributed queues • Applicable in many computations, e.g. Barnes-Hut, some graphics • Can compromise comm and locality, and increase synchronization • Whom to steal from, how many tasks to steal, ... • Termination detection Dynamic Tasking: • Maximum imbalance related to size of task • Deal with unpredictability in program or environment (e.g. All processes insert tasks P0 inserts P1 inserts P2 inserts P3 inserts Raytrace) – computation, communication, and memory system interactions –multippgrogramming and heterog eneity QQ 0 Q1 Q2 Q3 – used by runtime systems and OS too • Pool of tasks; take and add tasks until done • E.g. “self-scheduling” of loop iterations (shared loop counter) Others may steal All remove tasks P0 removes P1 removes P2 removes P3 removes (a) Centralized task queue (b) Distributed task queues (one per pr ocess) –11– CS 418 –12– CS 418 Page 3 Impact of Dynamic Assignment Load Balance and Synch Wait Time Sequential Work On SGI Origin 2000 (cache-coherent shared memory): Limit on speedup: Speedupproblem(p) < Max Work on any Processor Origin, semistatic Origg,yin, dynamic 30 30 Challenge, semistatic Challenge, dynamic • Work includes data access and other costs Origin, static Origin, static • Not just equal work, but must be busy at same time 25 Challenge, static 25 Challenge, static 20 20 Four parts to load balance and reducing synch wait time: 15 15 1. Identify enough concurrency Speedup Speedup 10 10 2. Decide how to manage it 5 5 3. Determine the granularity at which to exploit it 4. Reduce serialization and cost of synchronization (a)0 (b) 0 1 3 5 7 9 1113151719 21 23 25 27 29 31 1 3 5 7 9 1113151719212325272931 Number of processors Number of processors Barnes-Hut Raytrace –13– CS 418 –14– CS 418 Determining Task Granularity Load Balance and Synch Wait Time Sequential Work Task granularity: amount of work associated with a task Limit on speedup: Speedupproblem(p) < Max Work on any Processor • Work includes data access and other costs General rule: • Not just equal work, but must be busy at same time • Coarse-grained => often less load balance • Fine-grained => more overhead; often more communication & contention Four parts to load balance and reducing synch wait time: 1. Identify enough concurrency Communication & contention actually affected by assignment, not size 2. Decide how to manage it • Overhead by size itself too, particularly with task queues 3. Determine the granularity at which to exploit it 4. Reduce serialization and cost of synchronization –15– CS 418 –16– CS 418 Page 4 Reducing Serialization Partitioning for Performance Careful about assignment and orchestration (including 1. Balancing the workload and reducing wait time at synch scheduling) points EtEvent synchhitironization 2. Reducing inherent communication • Reduce use of conservative synchronization – e.g. point-to-point instead of barriers, or granularity of pt-to-pt 3. Reducing extra work • But fine-grained synch more difficult to program, more synch ops. Mutual exclusion • Separate locks for separate data – e.g. locking records in a database: lock per process, record, or field –lock ppqpqer task in task queue, not per queue – finer grain => less contention/serialization, more space, less reuse • Smaller, less frequent critical sections – don’t do reading/testing in critical section, only modification – e.g. searching for task to dequeue in task queue, building tree • Stagger critical sections in time –17– CS 418 –18– CS 418 Reducing Inherent Communication Domain Decomposition Communication is expensive! Works well for scientific, engineering, graphics, ... Measure: communication to computation ratio applications Focus here on inherent communication Exploits local-biased nature of physical problems • Determined by assignment of tasks to processes • Information requirements often short-range • Later see that actual communication can be greater • Or long-range but fall off with distance Assign tasks that access same data to same process Simple example: nearest-neighbor grid computation n n p Solving communication and load balance NP-hard in P0 P1 P2 P3 general case P P4 P5 P6 7 n n p But simple heuristic solutions work well in practice P8 P9 P10 P11 • Applications have structure! P12 P13 P14 P15 Perimeter to Area comm-to-comp ratio (area to volume in 3D) •Depends on n,p: decreases with n, increases with p –19– CS 418 –20– CS 418 Page 5 Domain Decomposition (Continued) Finding a Domain Decomposition Best domain decomposition depends on information requirements Static, by inspection Nearest neighbor example: block versus strip decomposition • Must be predictable: grid example above, and Ocean -----n- p n Static, but not by inspection • Input-dependent, require analyzing input structure P0 P1 P2 P3 • E.g sparse matrix computations, data mining P4 P5 P6 P7 n ------ n Semi-static (periodic repartitioning) p P8 P9 P10 P11 • Characteristics change but slowly; e.g. Barnes-Hut P12 P13 P14 P15 Static or semi-static, with dynamic task stealing • In itia l decompos ition, bu t hig hly unpredic ta ble; e.g ray ttiracing 4*√p 2*p Comm to comp: n for block, n for strip • Retain block from here on Application dependent: strip may be better in other cases • E.g. particle flow in tunnel –21– CS 418 –22– CS 418 Other Techniques Implications of Comm-to-Comp Ratio Scatter Decomposition, e.g. initial partition in Raytrace If denominator is execution time, ratio gives average BW needs 1 2 1 2 1 2 1 2 If operation count, gives extremes in impact of latency and 34 34 34 34 1 2 bandwidth 1 2 1 2 1 2 1 2 • Latency: assume no latency hiding 34 34 34 34 • Bandwidth: assume all latency hidden 1 2 1 2 1 2 1 2 • Reality is somewhere in between 34 34 34 34 3 4 Actual impact of comm.
Load more

Recommended publications
  • Data-Level Parallelism
    Fall 2015 :: CSE 610 – Parallel Computer Architectures Data-Level Parallelism Nima Honarmand Fall 2015 :: CSE 610 – Parallel Computer Architectures Overview • Data Parallelism vs. Control Parallelism – Data Parallelism: parallelism arises from executing essentially the same code on a large number of objects – Control Parallelism: parallelism arises from executing different threads of control concurrently • Hypothesis: applications that use massively parallel machines will mostly exploit data parallelism – Common in the Scientific Computing domain • DLP originally linked with SIMD machines; now SIMT is more common – SIMD: Single Instruction Multiple Data – SIMT: Single Instruction Multiple Threads Fall 2015 :: CSE 610 – Parallel Computer Architectures Overview • Many incarnations of DLP architectures over decades – Old vector processors • Cray processors: Cray-1, Cray-2, …, Cray X1 – SIMD extensions • Intel SSE and AVX units • Alpha Tarantula (didn’t see light of day ) – Old massively parallel computers • Connection Machines • MasPar machines – Modern GPUs • NVIDIA, AMD, Qualcomm, … • Focus of throughput rather than latency Vector Processors 4 SCALAR VECTOR (1 operation) (N operations) r1 r2 v1 v2 + + r3 v3 vector length add r3, r1, r2 vadd.vv v3, v1, v2 Scalar processors operate on single numbers (scalars) Vector processors operate on linear sequences of numbers (vectors) 6.888 Spring 2013 - Sanchez and Emer - L14 What’s in a Vector Processor? 5 A scalar processor (e.g. a MIPS processor) Scalar register file (32 registers) Scalar functional units (arithmetic, load/store, etc) A vector register file (a 2D register array) Each register is an array of elements E.g. 32 registers with 32 64-bit elements per register MVL = maximum vector length = max # of elements per register A set of vector functional units Integer, FP, load/store, etc Some times vector and scalar units are combined (share ALUs) 6.888 Spring 2013 - Sanchez and Emer - L14 Example of Simple Vector Processor 6 6.888 Spring 2013 - Sanchez and Emer - L14 Basic Vector ISA 7 Instr.
    [Show full text]
  • High Performance Computing Through Parallel and Distributed Processing
    Yadav S. et al., J. Harmoniz. Res. Eng., 2013, 1(2), 54-64 Journal Of Harmonized Research (JOHR) Journal Of Harmonized Research in Engineering 1(2), 2013, 54-64 ISSN 2347 – 7393 Original Research Article High Performance Computing through Parallel and Distributed Processing Shikha Yadav, Preeti Dhanda, Nisha Yadav Department of Computer Science and Engineering, Dronacharya College of Engineering, Khentawas, Farukhnagar, Gurgaon, India Abstract : There is a very high need of High Performance Computing (HPC) in many applications like space science to Artificial Intelligence. HPC shall be attained through Parallel and Distributed Computing. In this paper, Parallel and Distributed algorithms are discussed based on Parallel and Distributed Processors to achieve HPC. The Programming concepts like threads, fork and sockets are discussed with some simple examples for HPC. Keywords: High Performance Computing, Parallel and Distributed processing, Computer Architecture Introduction time to solve large problems like weather Computer Architecture and Programming play forecasting, Tsunami, Remote Sensing, a significant role for High Performance National calamities, Defence, Mineral computing (HPC) in large applications Space exploration, Finite-element, Cloud science to Artificial Intelligence. The Computing, and Expert Systems etc. The Algorithms are problem solving procedures Algorithms are Non-Recursive Algorithms, and later these algorithms transform in to Recursive Algorithms, Parallel Algorithms particular Programming language for HPC. and Distributed Algorithms. There is need to study algorithms for High The Algorithms must be supported the Performance Computing. These Algorithms Computer Architecture. The Computer are to be designed to computer in reasonable Architecture is characterized with Flynn’s Classification SISD, SIMD, MIMD, and For Correspondence: MISD. Most of the Computer Architectures preeti.dhanda01ATgmail.com are supported with SIMD (Single Instruction Received on: October 2013 Multiple Data Streams).
    [Show full text]
  • 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
    [Show full text]
  • Demystifying Parallel and Distributed Deep Learning: an In-Depth Concurrency Analysis
    1 Demystifying Parallel and Distributed Deep Learning: An In-Depth Concurrency Analysis TAL BEN-NUN and TORSTEN HOEFLER, ETH Zurich, Switzerland Deep Neural Networks (DNNs) are becoming an important tool in modern computing applications. Accelerating their training is a major challenge and techniques range from distributed algorithms to low-level circuit design. In this survey, we describe the problem from a theoretical perspective, followed by approaches for its parallelization. We present trends in DNN architectures and the resulting implications on parallelization strategies. We then review and model the different types of concurrency in DNNs: from the single operator, through parallelism in network inference and training, to distributed deep learning. We discuss asynchronous stochastic optimization, distributed system architectures, communication schemes, and neural architecture search. Based on those approaches, we extrapolate potential directions for parallelism in deep learning. CCS Concepts: • General and reference → Surveys and overviews; • Computing methodologies → Neural networks; Parallel computing methodologies; Distributed computing methodologies; Additional Key Words and Phrases: Deep Learning, Distributed Computing, Parallel Algorithms ACM Reference Format: Tal Ben-Nun and Torsten Hoefler. 2018. Demystifying Parallel and Distributed Deep Learning: An In-Depth Concurrency Analysis. 47 pages. 1 INTRODUCTION Machine Learning, and in particular Deep Learning [143], is rapidly taking over a variety of aspects in our daily lives.
    [Show full text]
  • Exploiting Parallelism in Many-Core Architectures: Architectures G Bortolotti, M Caberletti, G Crimi Et Al
    Journal of Physics: Conference Series OPEN ACCESS Related content - Computing on Knights and Kepler Exploiting parallelism in many-core architectures: Architectures G Bortolotti, M Caberletti, G Crimi et al. Lattice Boltzmann models as a test case - Implementing the lattice Boltzmann model on commodity graphics hardware Arie Kaufman, Zhe Fan and Kaloian To cite this article: F Mantovani et al 2013 J. Phys.: Conf. Ser. 454 012015 Petkov - Many-core experience with HEP software at CERN openlab Sverre Jarp, Alfio Lazzaro, Julien Leduc et al. View the article online for updates and enhancements. Recent citations - Lattice Boltzmann benchmark kernels as a testbed for performance analysis M. Wittmann et al - Enrico Calore et al - Computing on Knights and Kepler Architectures G Bortolotti et al This content was downloaded from IP address 170.106.202.58 on 23/09/2021 at 15:57 24th IUPAP Conference on Computational Physics (IUPAP-CCP 2012) IOP Publishing Journal of Physics: Conference Series 454 (2013) 012015 doi:10.1088/1742-6596/454/1/012015 Exploiting parallelism in many-core architectures: Lattice Boltzmann models as a test case F Mantovani1, M Pivanti2, S F Schifano3 and R Tripiccione4 1 Department of Physics, Univesit¨atRegensburg, Germany 2 Department of Physics, Universit`adi Roma La Sapienza, Italy 3 Department of Mathematics and Informatics, Universit`adi Ferrara and INFN, Italy 4 Department of Physics and CMCS, Universit`adi Ferrara and INFN, Italy E-mail: [email protected], [email protected], [email protected], [email protected] Abstract. In this paper we address the problem of identifying and exploiting techniques that optimize the performance of large scale scientific codes on many-core processors.
    [Show full text]
  • An Introduction to Gpus, CUDA and Opencl
    An Introduction to GPUs, CUDA and OpenCL Bryan Catanzaro, NVIDIA Research Overview ¡ Heterogeneous parallel computing ¡ The CUDA and OpenCL programming models ¡ Writing efficient CUDA code ¡ Thrust: making CUDA C++ productive 2/54 Heterogeneous Parallel Computing Latency-Optimized Throughput- CPU Optimized GPU Fast Serial Scalable Parallel Processing Processing 3/54 Why do we need heterogeneity? ¡ Why not just use latency optimized processors? § Once you decide to go parallel, why not go all the way § And reap more benefits ¡ For many applications, throughput optimized processors are more efficient: faster and use less power § Advantages can be fairly significant 4/54 Why Heterogeneity? ¡ Different goals produce different designs § Throughput optimized: assume work load is highly parallel § Latency optimized: assume work load is mostly sequential ¡ To minimize latency eXperienced by 1 thread: § lots of big on-chip caches § sophisticated control ¡ To maXimize throughput of all threads: § multithreading can hide latency … so skip the big caches § simpler control, cost amortized over ALUs via SIMD 5/54 Latency vs. Throughput Specificaons Westmere-EP Fermi (Tesla C2050) 6 cores, 2 issue, 14 SMs, 2 issue, 16 Processing Elements 4 way SIMD way SIMD @3.46 GHz @1.15 GHz 6 cores, 2 threads, 4 14 SMs, 48 SIMD Resident Strands/ way SIMD: vectors, 32 way Westmere-EP (32nm) Threads (max) SIMD: 48 strands 21504 threads SP GFLOP/s 166 1030 Memory Bandwidth 32 GB/s 144 GB/s Register File ~6 kB 1.75 MB Local Store/L1 Cache 192 kB 896 kB L2 Cache 1.5 MB 0.75 MB
    [Show full text]
  • 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]>
    [Show full text]
  • Computer Architecture: Parallel Processing Basics
    Computer Architecture: Parallel Processing Basics Onur Mutlu & Seth Copen Goldstein Carnegie Mellon University 9/9/13 Today What is Parallel Processing? Why? Kinds of Parallel Processing Multiprocessing and Multithreading Measuring success Speedup Amdhal’s Law Bottlenecks to parallelism 2 Concurrent Systems Embedded-Physical Distributed Sensor Claytronics Networks Concurrent Systems Embedded-Physical Distributed Sensor Claytronics Networks Geographically Distributed Power Internet Grid Concurrent Systems Embedded-Physical Distributed Sensor Claytronics Networks Geographically Distributed Power Internet Grid Cloud Computing EC2 Tashi PDL'09 © 2007-9 Goldstein5 Concurrent Systems Embedded-Physical Distributed Sensor Claytronics Networks Geographically Distributed Power Internet Grid Cloud Computing EC2 Tashi Parallel PDL'09 © 2007-9 Goldstein6 Concurrent Systems Physical Geographical Cloud Parallel Geophysical +++ ++ --- --- location Relative +++ +++ + - location Faults ++++ +++ ++++ -- Number of +++ +++ + - Processors + Network varies varies fixed fixed structure Network --- --- + + connectivity 7 Concurrent System Challenge: Programming The old joke: How long does it take to write a parallel program? One Graduate Student Year 8 Parallel Programming Again?? Increased demand (multicore) Increased scale (cloud) Improved compute/communicate Change in Application focus Irregular Recursive data structures PDL'09 © 2007-9 Goldstein9 Why Parallel Computers? Parallelism: Doing multiple things at a time Things: instructions,
    [Show full text]
  • 1 Parallelism
    Lecture 23: Parallelism • Administration – Take QUIZ 17 over P&H 7.1-5, before 11:59pm today – Project: Cache Simulator, Due April 29, 2010 • Last Time – On chip communication – How I/O works • Today – Where do architectures exploit parallelism? – What are the implications for programming models? – What are the implications for communication? – … for caching? UTCS 352, Lecture 23 1 Parallelism • What type of parallelism do applications have? • How can computer architectures exploit this parallelism? UTCS 352, Lecture 23 2 1 Granularity of Parallelism • Fine grain instruction level parallelism • Fine grain data parallelism • Coarse grain (data center) parallelism • Multicore parallelism UTCS 352, Lecture 23 3 Fine grain instruction level parallelism • Fine grain instruction level parallelism – Pipelining – Multi-issue (dynamic scheduling & issue) – VLIW (static issue) – Very Long Instruction Word • Each VLIW instruction includes up to N RISC-like operations • Compiler groups up to N independent operations • Architecture issues fixed size VLIW instructions UTCS 352, Lecture 23 4 2 VLIW Example Theory Let N = 4 Theoretically 4 ops/cycle • 8 VLIW = 32 operations in practice In practice, • Compiler cannot fill slots • Memory stalls - load stall - UTCS 352, Lecture 23 5 Multithreading • Performing multiple threads together – Replicate registers, PC, etc. – Fast switching between threads • Fine-grain multithreading – Switch threads after each cycle – Interleave instruction execution – If one thread stalls, others are executed • Medium-grain
    [Show full text]
  • Exploiting Parallelism Opportunities with Deep Learning Frameworks
    Exploiting Parallelism Opportunities with Deep Learning Frameworks YU EMMA WANG, Harvard University CAROLE-JEAN WU, Facebook XIAODONG WANG, Facebook KIM HAZELWOOD, Facebook DAVID BROOKS, Harvard University State-of-the-art machine learning frameworks support a wide variety of design features to enable a flexible machine learning programming interface and to ease the programmability burden on machine learning developers. Identifying and using a performance-optimal setting in feature-rich frameworks, however, involves a non-trivial amount of performance profiling efforts and often relies on domain-specific knowledge. This paper takesa deep dive into analyzing the performance impact of key design features in a machine learning framework and quantifies the role of parallelism. The observations and insights distill into a simple set of guidelines thatone can use to achieve much higher training and inference speedup. Across a diverse set of real-world deep learning models, the evaluation results show that the proposed performance tuning guidelines outperform the Intel and TensorFlow recommended settings by 1.29× and 1.34×, respectively. ACM Reference Format: Yu Emma Wang, Carole-Jean Wu, Xiaodong Wang, Kim Hazelwood, and David Brooks. 2020. Exploiting Parallelism Opportunities with Deep Learning Frameworks. 1, 1 (July 2020), 21 pages. https://doi.org/10.1145/ nnnnnnn.nnnnnnn 1 INTRODUCTION The wide adoption of deep learning (DL) tasks has spawned a plethora of domain-specific frameworks to help improve development efficiency and to enable a flexible path from research to production deployment for machine learning (ML). Notable ML frameworks include Caffe/Caffe2 [23], Py- Torch [24], TensorFlow [2], and MXNet [8]. These frameworks provide high-level APIs as the building blocks of DL models.
    [Show full text]
  • An Introduction to CUDA/Opencl and Graphics Processors
    An Introduction to CUDA/OpenCL and Graphics Processors Bryan Catanzaro, NVIDIA Research Overview ¡ Terminology ¡ The CUDA and OpenCL programming models ¡ Understanding how CUDA maps to NVIDIA GPUs ¡ Thrust 2/74 Heterogeneous Parallel Computing Latency Throughput Optimized CPU Optimized GPU Fast Serial Scalable Parallel Processing Processing 3/74 Latency vs. Throughput Latency Throughput ¡ Latency: yoke of oxen § Each core optimized for executing a single thread ¡ Throughput: flock of chickens § Cores optimized for aggregate throughput, deemphasizing individual performance ¡ (apologies to Seymour Cray) 4/74 Latency vs. Throughput, cont. Specificaons Sandy Bridge- Kepler EP (Tesla K20) 8 cores, 2 issue, 14 SMs, 6 issue, 32 Processing Elements 8 way SIMD way SIMD @3.1 GHz @730 MHz 8 cores, 2 threads, 8 14 SMs, 64 SIMD Resident Strands/ way SIMD: vectors, 32 way Threads (max) SIMD: Sandy Bridge-EP (32nm) 96 strands 28672 threads SP GFLOP/s 396 3924 Memory Bandwidth 51 GB/s 250 GB/s Register File 128 kB (?) 3.5 MB Local Store/L1 Cache 256 kB 896 kB L2 Cache 2 MB 1.5 MB L3 Cache 20 MB - Kepler (28nm) 5/74 Why Heterogeneity? ¡ Different goals produce different designs § Manycore assumes work load is highly parallel § Multicore must be good at everything, parallel or not ¡ Multicore: minimize latency experienced by 1 thread § lots of big on-chip caches § extremely sophisticated control ¡ Manycore: maximize throughput of all threads § lots of big ALUs § multithreading can hide latency … so skip the big caches § simpler control, cost amortized over
    [Show full text]
  • Introduction to Parallel Machines and Programming Models Lecture 3
    Outline •"Overview of parallel machines (~hardware) and CS 267: programming models (~software) •" Shared memory Introduction to Parallel Machines •" Shared address space and Programming Models •" Message passing Lecture 3 •" Data parallel " •" Clusters of SMPs or GPUs •" Grid James Demmel •"Note: Parallel machine may or may not be tightly www.cs.berkeley.edu/~demmel/cs267_Spr15/! coupled to programming model ! •" Historically, tight coupling ! •" Today, portability is important CS267 Lecture 3! 1! 01/27/2015 CS267 Lecture 3! 2! 01/28/2011! A generic parallel architecture Parallel Programming Models •" Programming model is made up of the languages and libraries that create an abstract view of the machine Proc Proc Proc Proc Proc •" Control Proc •" How is parallelism created? •" What orderings exist between operations? Interconnection Network •" Data •" What data is private vs. shared? •" How is logically shared data accessed or communicated? Memory Memory Memory Memory Memory •" Synchronization •" What operations can be used to coordinate parallelism? •" What are the atomic (indivisible) operations? •" Where is the memory physically located? •" Cost •" Is it connected directly to processors? •" How do we account for the cost of each of the above? •" What is the connectivity of the network? 01/27/2015 CS267 Lecture 3! 3! 01/27/2015 CS267 Lecture 3! 4! CS267 Lecture 2 1 Simple Example Programming Model 1: Shared Memory •"Consider applying a function f to the elements •" Program is a collection of threads of control. of an array A and then computing its sum: •" Can be created dynamically, mid-execution, in some languages n−1 •" Each thread has a set of private variables, e.g., local stack variables ∑ f (A[i]) •" Also a set of shared variables, e.g., static variables, shared common •"Questions: i=0 blocks, or global heap.
    [Show full text]