DOCSLIB.ORG

GPU Computing with Apache Spark and Python

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
GPU Computing with Apache Spark and Python GPU Computing with Apache Spark and Python Stan Seibert Siu Kwan Lam April 5, 2016 © 2015 Continuum Analytics- Confidential & Proprietary My Background • Trained in particle physics • Using Python for data analysis for 10 years • Using GPUs for data analysis for 7 years • Currently lead the High Performance Python team at Continuum 2 About Continuum Analytics • We give superpowers to people who change the world! • We want to help everyone analyze their data with Python (and other tools), so we offer: • Enterprise Products • Consulting • Training • Open Source 3 • I’m going to use Anaconda throughout this presentation. • Anaconda is a free Mac/Win/Linux Python distribution: • Based on conda, an open source package manager • Installs both Python and non-Python dependencies • Easiest way to get the software I will talk about today • https://www.continuum.io/downloads 4 Overview 1. Why Python? 2. Numba: A Python JIT compiler for the CPU and GPU 3. PySpark: Distributed computing for Python 4. Example: Image Registration 5. Tips and Tricks 6. Conclusion WHY PYTHON? 6 Why is Python so popular? • Straightforward, productive language for system administrators, programmers, scientists, analysts and hobbyists • Great community: • Lots of tutorial and reference materials • Easy to interface with other languages • Vast ecosystem of useful libraries 7 8 … But, isn’t Python slow? • Pure, interpreted Python is slow. • Python excels at interfacing with other languages used in HPC: C: ctypes, CFFI, Cython C++: Cython, Boost.Python FORTRAN: f2py • Secret: Most scientific Python packages put the speed critical sections of their algorithms in a compiled language. 9 Is there another way? • Switching languages for speed in your projects can be a little clunky: • Sometimes tedious boilerplate for translating data types across the language barrier • Generating compiled functions for the wide range of data types can be difficult • How can we use cutting edge hardware, like GPUs? 10 NUMBA: A PYTHON JIT COMPILER 11 Compiling Python • Numba is an open-source, type-specializing compiler for Python functions • Can translate Python syntax into machine code if all type information can be deduced when the function is called. • Implemented as a module. Does not replace the Python interpreter! • Code generation done with: • LLVM (for CPU) • NVVM (for CUDA GPUs). 12 Supported Platforms OS HW SW • Windows (7 and later) • 32 and 64-bit x86 CPUs • Python 2 and 3 • OS X (10.9 and later) • CUDA-capable NVIDIA GPUs • NumPy 1.7 through 1.10 • Linux (~RHEL 5 and • HSA-capable AMD GPUs later) • Experimental support for ARMv7 (Raspberry Pi 2) 13 How Does Numba Work? @jit def do_math(a, b): … >>> do_math(x, y) Python Function Functions (bytecode) Arguments Type Rewrite IR Inference Bytecode Numba IR Analysis Lowering Cache Execute! Machine LLVM/NVVM JIT LLVM IR Code 14 Numba on the CPU 15 Numba on the CPU Numba decorator (nopython=True not required) Array Allocation Looping over ndarray x as an iterator Using numpy math functions Returning a slice of the array 2.7x speedup! 16 CUDA Kernels in Python 17 CUDA Kernels in Python Decorator will infer type signature when you call it Helper function to compute blockIdx.x * blockDim.x + threadIdx.x NumPy arrays have expected attributes and indexing Helper function to compute blockDim.x * gridDim.x 18 Calling the Kernel from Python Works just like CUDA C, except Numba handles allocating and copying data to/from the host if needed 19 Handling Device Memory Directly Memory allocation matters in small tasks. 20 Higher Level Tools: GPU Ufuncs 21 Higher Level Tools: GPU Ufuncs Decorator for creating ufunc List of supported type signatures Code generation target 22 GPU Ufuncs Performance 4x speedup incl. host<->device round trip on GeForce GT 650M 23 Accelerate Library Bindings: cuFFT MKL accelerated FFT >2x speedup incl. host<->device round trip on GeForce GT 650M 24 PYSPARK: DISTRIBUTED COMPUTING FOR PYTHON 25 What is Apache Spark? • An API and an execution engine for distributed computing on a cluster • Based on the concept of Resilient Distributed Datasets (RDDs) • Dataset: Collection of independent elements (files, objects, etc) in memory from previous calculations, or originating from some data store • Distributed: Elements in RDDs are grouped into partitions and may be stored on different nodes • Resilient: RDDs remember how they were created, so if a node goes down, Spark can recompute the lost elements on another node 26 Computation DAGs Fig from: https://databricks.com/blog/2015/06/22/understanding-your-spark-application-through-visualization.html 27 How does Spark Scale? • All cluster scaling is about minimizing I/O. Spark does this in several ways: • Keep intermediate results in memory with rdd.cache() • Move computation to the data whenever possible (functions are small and data is big!) • Provide computation primitives that expose parallelism and minimize communication between workers: map, filter, sample, reduce, … 28 Python and Spark • Spark is implemented in Java & Scala on the JVM • Full API support for Scala, Java, and Python (+ limited support for R) • How does Python work, since it doesn’t run on the JVM? (not counting IronPython) 29 Spark Py4J Python Worker Interpreter Java Python Spark Py4J Spark Context Context Spark Py4J Python Python Java Worker Interpreter Java Python Client Cluster 30 Getting Started • conda can create a local environment for Spark for you: conda create -n spark -c anaconda-cluster python=3.5 spark numba \ cudatoolkit ipython-notebook source activate spark #Uncomment below if on Mac OS X #export JRE_HOME=$(/usr/libexec/java_home) #export JAVA_HOME=$(/usr/libexec/java_home) IPYTHON_OPTS="notebook" pyspark # starts jupyter notebook 31 Using Numba (CPU) with Spark 32 Using Numba (CPU) with Spark Load arrays onto the Spark Workers Apply my function to every element in the RDD and return first element 33 Using CUDA Python with Spark Define CUDA kernel Compilation happens here Wrap CUDA kernel launching logic Creates Spark RDD (8 partitions) Apply gpu_work on each partition 34 recompile if CUDA arch is different LLVM IR LLVM IR PTX CUDA Python Compiles Serialize Deserialize Finalize CUDA Kernel Binary PTX PTX use serialized PTX if CUDA arch matches client Client Cluster 35 LLVM IR LLVM IR PTX CUDA Python Compiles Serialize Deserialize Finalize CUDA Kernel Binary PTX PTX This happens on every worker process Client Cluster 36 EXAMPLE: IMAGE REGISTRATION 37 Basic Algorithm image set Group similar images (unsupervised kNN clustering) attempt image registration on every pair in each group (phase correlation based; FFT heavy) unused images new images Progress? 38 Basic Algorithm • Reduce number of pairwise image set image registration attempt • Run on CPU Group similar images (unsupervised kNN clustering) attempt image registration on every pair in each group • FFT 2D heavy (phase correlation based; FFT heavy) • Expensive • Mostly running on GPU unused images new images Progress? 39 Cross Power Spectrum Core of phase correlation based image registration algorithm 40 Cross Power Spectrum Core of phase correlation based image registration algorithm cuFFT explicit memory transfer 41 Scaling on Spark Image set • Machines have multiple GPUs • Each worker computes a partition at a time Random partitioning rdd.repartition(num_parts) partition partition partition Apply ImgReg on each Partition i i i rdd.mapPartitions(imgreg_func) … u n u n u n … Progress? 42 CUDA Multi-Process Service (MPS) • Sharing one GPU between multiple workers can be beneficial • nvidia-cuda-mps-control • Better GPU utilization from multiple processes • For our example app: 5-10% improvement with MPS • Effect can be bigger for app with higher GPU utilization 43 Scaling with Multiple GPUs and MPS • 1 CPU worker: 1 Tesla K20 • 2 CPU worker: 2 Tesla K20 • 4 CPU worker: 2 Tesla K20 (2 workers per GPU using MPS) • Spark + Dask = External process performing GPU calculation 44 Spark vs CUDA • Spark logic is similar to CUDA host logic • Partitions are like CUDA blocks • mapPartitions is like kernel launch • Work cooperatively within a partition • Spark network transfer overhead vs PCI-express transfer overhead Image set Random partitioning partition partition partition rdd.repartition(num_parts i i i … Apply ImgReg on each Partition u n u n u n … rdd.mapPartitions(imgreg_func) Progress? 45 TIPS AND TRICKS 46 Parallelism is about communication • Be aware of the communication overhead when deciding how to chunk your work in Spark. • Bigger chunks: Risk of underutilization of resources • Smaller chunks: Risk of computation swamped by overhead ➡ Start with big chunks, then move to smaller ones if you fail to reach full utilization. 47 Start Small • Easier to debug your logic on your local system! • Run a Spark cluster on your workstation where it is easier to debug and profile. • Move to the cluster only after you’ve seen things work locally. (Spark makes this fairly painless.) 48 Be interactive with Big Data! • Run Jupyter notebook on Spark + MultiGPU cluster • Live experimentation helps you develop and understand your progress • Spark keeps track of intermediate data and recomputes them if they are lost 49 Amortize Setup Costs • Make sure that one-time costs are actually done once: • GPU state (like FFT plans) can be slow to initialize. Do this at the top of your mapPartition call and reuse for each element in the RDD partition. • Coarser chunking of tasks allows GPU memory allocations to be reused between steps inside a single task. 50 Be(a)ware
Load more

Recommended publications
  • Python on Gpus (Work in Progress!)
    Python on GPUs (work in progress!) Laurie Stephey GPUs for Science Day, July 3, 2019 Rollin Thomas, NERSC Lawrence Berkeley National Laboratory Python is friendly and popular Screenshots from: https://www.tiobe.com/tiobe-index/ So you want to run Python on a GPU? You have some Python code you like. Can you just run it on a GPU? import numpy as np from scipy import special import gpu ? Unfortunately no. What are your options? Right now, there is no “right” answer ● CuPy ● Numba ● pyCUDA (https://mathema.tician.de/software/pycuda/) ● pyOpenCL (https://mathema.tician.de/software/pyopencl/) ● Rewrite kernels in C, Fortran, CUDA... DESI: Our case study Now Perlmutter 2020 Goal: High quality output spectra Spectral Extraction CuPy (https://cupy.chainer.org/) ● Developed by Chainer, supported in RAPIDS ● Meant to be a drop-in replacement for NumPy ● Some, but not all, NumPy coverage import numpy as np import cupy as cp cpu_ans = np.abs(data) #same thing on gpu gpu_data = cp.asarray(data) gpu_temp = cp.abs(gpu_data) gpu_ans = cp.asnumpy(gpu_temp) Screenshot from: https://docs-cupy.chainer.org/en/stable/reference/comparison.html eigh in CuPy ● Important function for DESI ● Compared CuPy eigh on Cori Volta GPU to Cori Haswell and Cori KNL ● Tried “divide-and-conquer” approach on both CPU and GPU (1, 2, 5, 10 divisions) ● Volta wins only at very large matrix sizes ● Major pro: eigh really easy to use in CuPy! legval in CuPy ● Easy to convert from NumPy arrays to CuPy arrays ● This function is ~150x slower than the cpu version! ● This implies there
    [Show full text]
  • Introduction Shrinkage Factor Reference
    Comparison study for implementation efficiency of CUDA GPU parallel computation with the fast iterative shrinkage-thresholding algorithm Younsang Cho, Donghyeon Yu Department of Statistics, Inha university 4. TensorFlow functions in Python (TF-F) Introduction There are some functions executed on GPU in TensorFlow. So, we implemented our algorithm • Parallel computation using graphics processing units (GPUs) gets much attention and is just using that functions. efficient for single-instruction multiple-data (SIMD) processing. 5. Neural network with TensorFlow in Python (TF-NN) • Theoretical computation capacity of the GPU device has been growing fast and is much higher Neural network model is flexible, and the LASSO problem can be represented as a simple than that of the CPU nowadays (Figure 1). neural network with an ℓ1-regularized loss function • There are several platforms for conducting parallel computation on GPUs using compute 6. Using dynamic link library in Python (P-DLL) unified device architecture (CUDA) developed by NVIDIA. (Python, PyCUDA, Tensorflow, etc. ) As mentioned before, we can load DLL files, which are written in CUDA C, using "ctypes.CDLL" • However, it is unclear what platform is the most efficient for CUDA. that is a built-in function in Python. 7. Using dynamic link library in R (R-DLL) We can also load DLL files, which are written in CUDA C, using "dyn.load" in R. FISTA (Fast Iterative Shrinkage-Thresholding Algorithm) We consider FISTA (Beck and Teboulle, 2009) with backtracking as the following: " Step 0. Take �! > 0, some � > 1, and �! ∈ ℝ . Set �# = �!, �# = 1. %! Step k. � ≥ 1 Find the smallest nonnegative integers �$ such that with �g = � �$&# � �(' �$ ≤ �(' �(' �$ , �$ .
    [Show full text]
  • Prototyping and Developing GPU-Accelerated Solutions with Python and CUDA Luciano Martins and Robert Sohigian, 2018-11-22 Introduction to Python
    Prototyping and Developing GPU-Accelerated Solutions with Python and CUDA Luciano Martins and Robert Sohigian, 2018-11-22 Introduction to Python GPU-Accelerated Computing NVIDIA® CUDA® technology Why Use Python with GPUs? Agenda Methods: PyCUDA, Numba, CuPy, and scikit-cuda Summary Q&A 2 Introduction to Python Released by Guido van Rossum in 1991 The Zen of Python: Beautiful is better than ugly. Explicit is better than implicit. Simple is better than complex. Complex is better than complicated. Flat is better than nested. Interpreted language (CPython, Jython, ...) Dynamically typed; based on objects 3 Introduction to Python Small core structure: ~30 keywords ~ 80 built-in functions Indentation is a pretty serious thing Dynamically typed; based on objects Binds to many different languages Supports GPU acceleration via modules 4 Introduction to Python 5 Introduction to Python 6 Introduction to Python 7 GPU-Accelerated Computing “[T]the use of a graphics processing unit (GPU) together with a CPU to accelerate deep learning, analytics, and engineering applications” (NVIDIA) Most common GPU-accelerated operations: Large vector/matrix operations (Basic Linear Algebra Subprograms - BLAS) Speech recognition Computer vision 8 GPU-Accelerated Computing Important concepts for GPU-accelerated computing: Host ― the machine running the workload (CPU) Device ― the GPUs inside of a host Kernel ― the code part that runs on the GPU SIMT ― Single Instruction Multiple Threads 9 GPU-Accelerated Computing 10 GPU-Accelerated Computing 11 CUDA Parallel computing
    [Show full text]
  • Tangent: Automatic Differentiation Using Source-Code Transformation for Dynamically Typed Array Programming
    Tangent: Automatic differentiation using source-code transformation for dynamically typed array programming Bart van Merriënboer Dan Moldovan Alexander B Wiltschko MILA, Google Brain Google Brain Google Brain [email protected] [email protected] [email protected] Abstract The need to efficiently calculate first- and higher-order derivatives of increasingly complex models expressed in Python has stressed or exceeded the capabilities of available tools. In this work, we explore techniques from the field of automatic differentiation (AD) that can give researchers expressive power, performance and strong usability. These include source-code transformation (SCT), flexible gradient surgery, efficient in-place array operations, and higher-order derivatives. We implement and demonstrate these ideas in the Tangent software library for Python, the first AD framework for a dynamic language that uses SCT. 1 Introduction Many applications in machine learning rely on gradient-based optimization, or at least the efficient calculation of derivatives of models expressed as computer programs. Researchers have a wide variety of tools from which they can choose, particularly if they are using the Python language [21, 16, 24, 2, 1]. These tools can generally be characterized as trading off research or production use cases, and can be divided along these lines by whether they implement automatic differentiation using operator overloading (OO) or SCT. SCT affords more opportunities for whole-program optimization, while OO makes it easier to support convenient syntax in Python, like data-dependent control flow, or advanced features such as custom partial derivatives. We show here that it is possible to offer the programming flexibility usually thought to be exclusive to OO-based tools in an SCT framework.
    [Show full text]
  • Python Guide Documentation 0.0.1
    Python Guide Documentation 0.0.1 Kenneth Reitz 2015 11 07 Contents 1 3 1.1......................................................3 1.2 Python..................................................5 1.3 Mac OS XPython.............................................5 1.4 WindowsPython.............................................6 1.5 LinuxPython...............................................8 2 9 2.1......................................................9 2.2...................................................... 15 2.3...................................................... 24 2.4...................................................... 25 2.5...................................................... 27 2.6 Logging.................................................. 31 2.7...................................................... 34 2.8...................................................... 37 3 / 39 3.1...................................................... 39 3.2 Web................................................... 40 3.3 HTML.................................................. 47 3.4...................................................... 48 3.5 GUI.................................................... 49 3.6...................................................... 51 3.7...................................................... 52 3.8...................................................... 53 3.9...................................................... 58 3.10...................................................... 59 3.11...................................................... 62
    [Show full text]
  • Hyperlearn Documentation Release 1
    HyperLearn Documentation Release 1 Daniel Han-Chen Jun 19, 2020 Contents 1 Example code 3 1.1 hyperlearn................................................3 1.1.1 hyperlearn package.......................................3 1.1.1.1 Submodules......................................3 1.1.1.2 hyperlearn.base module................................3 1.1.1.3 hyperlearn.linalg module...............................7 1.1.1.4 hyperlearn.utils module................................ 10 1.1.1.5 hyperlearn.random module.............................. 11 1.1.1.6 hyperlearn.exceptions module............................. 11 1.1.1.7 hyperlearn.multiprocessing module.......................... 11 1.1.1.8 hyperlearn.numba module............................... 11 1.1.1.9 hyperlearn.solvers module............................... 12 1.1.1.10 hyperlearn.stats module................................ 14 1.1.1.11 hyperlearn.big_data.base module........................... 15 1.1.1.12 hyperlearn.big_data.incremental module....................... 15 1.1.1.13 hyperlearn.big_data.lsmr module........................... 15 1.1.1.14 hyperlearn.big_data.randomized module....................... 16 1.1.1.15 hyperlearn.big_data.truncated module........................ 17 1.1.1.16 hyperlearn.decomposition.base module........................ 18 1.1.1.17 hyperlearn.decomposition.NMF module....................... 18 1.1.1.18 hyperlearn.decomposition.PCA module........................ 18 1.1.1.19 hyperlearn.decomposition.PCA module........................ 18 1.1.1.20 hyperlearn.discriminant_analysis.base
    [Show full text]
  • Zmcintegral: a Package for Multi-Dimensional Monte Carlo Integration on Multi-Gpus
    ZMCintegral: a Package for Multi-Dimensional Monte Carlo Integration on Multi-GPUs Hong-Zhong Wua,∗, Jun-Jie Zhanga,∗, Long-Gang Pangb,c, Qun Wanga aDepartment of Modern Physics, University of Science and Technology of China bPhysics Department, University of California, Berkeley, CA 94720, USA and cNuclear Science Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA Abstract We have developed a Python package ZMCintegral for multi-dimensional Monte Carlo integration on multiple Graphics Processing Units(GPUs). The package employs a stratified sampling and heuristic tree search algorithm. We have built three versions of this package: one with Tensorflow and other two with Numba, and both support general user defined functions with a user- friendly interface. We have demonstrated that Tensorflow and Numba help inexperienced scientific researchers to parallelize their programs on multiple GPUs with little work. The precision and speed of our package is compared with that of VEGAS for two typical integrands, a 6-dimensional oscillating function and a 9-dimensional Gaussian function. The results show that the speed of ZMCintegral is comparable to that of the VEGAS with a given precision. For heavy calculations, the algorithm can be scaled on distributed clusters of GPUs. Keywords: Monte Carlo integration; Stratified sampling; Heuristic tree search; Tensorflow; Numba; Ray. PROGRAM SUMMARY Manuscript Title: ZMCintegral: a Package for Multi-Dimensional Monte Carlo Integration on Multi-GPUs arXiv:1902.07916v2 [physics.comp-ph] 2 Oct 2019 Authors: Hong-Zhong Wu; Jun-Jie Zhang; Long-Gang Pang; Qun Wang Program Title: ZMCintegral Journal Reference: ∗Both authors contributed equally to this manuscript. Preprint submitted to Computer Physics Communications October 3, 2019 Catalogue identifier: Licensing provisions: Apache License Version, 2.0(Apache-2.0) Programming language: Python Operating system: Linux Keywords: Monte Carlo integration; Stratified sampling; Heuristic tree search; Tensorflow; Numba; Ray.
    [Show full text]
  • Python Guide Documentation Publicación 0.0.1
    Python Guide Documentation Publicación 0.0.1 Kenneth Reitz 17 de May de 2018 Índice general 1. Empezando con Python 3 1.1. Eligiendo un Interprete Python (3 vs. 2).................................3 1.2. Instalando Python Correctamente....................................5 1.3. Instalando Python 3 en Mac OS X....................................6 1.4. Instalando Python 3 en Windows....................................8 1.5. Instalando Python 3 en Linux......................................9 1.6. Installing Python 2 on Mac OS X.................................... 10 1.7. Instalando Python 2 en Windows.................................... 12 1.8. Installing Python 2 on Linux....................................... 13 1.9. Pipenv & Ambientes Virtuales...................................... 14 1.10. Un nivel más bajo: virtualenv...................................... 17 2. Ambientes de Desarrollo de Python 21 2.1. Your Development Environment..................................... 21 2.2. Further Configuration of Pip and Virtualenv............................... 26 3. Escribiendo Buen Código Python 29 3.1. Estructurando tu Proyecto........................................ 29 3.2. Code Style................................................ 40 3.3. Reading Great Code........................................... 49 3.4. Documentation.............................................. 50 3.5. Testing Your Code............................................ 53 3.6. Logging.................................................. 57 3.7. Common Gotchas...........................................
    [Show full text]
  • Aayush Mandhyan
    AAYUSH MANDHYAN New Brunswick, New Jersey | 848-218-8606 | LinkedIn | GitHub | Portfolio | [email protected] Education Rutgers University, New Brunswick September 2018 – May 2020 • Master of Science in Data Science, CGPA – 3.75/4 • Relevant Coursework: Machine Learning, Reinforcement Learning, Introduction to Artificial Intelligence, Data Interaction and Visual Analytics, Massive Data Storage and Retrieval, Probability and Statistical Inference. SRM University, NCR Campus, Ghaziabad, India August 2012 – May 2016 • Bachelor of technology in Computer Science and Engineering, CGPA – 8/10 Skills • Algorithms: Q-Learning, Neural Networks, LSTM, RNN, CNN, Auto-Encoders, XGBoost, SVM, Random Forest, Decision Trees, Logistic Regression, Lasso Regression, Ridge Regression, KNN, etc. • Languages: Python, R, Java, HTML, JavaScript • Libraries: TensorFlow, PyTorch, Keras, Scikit-learn, XGBoost, NumPy, Pandas, Matplotlib, CuPy, Numba, OpenCV, PySpark, NLTK, Gensim, Flask, R Shiny. • Tools: MySQL, NoSQL, MongoDB, REST API, Linux, Git, Jupyter, AWS, GCP, Openstack, Docker. • Technologies: Deep Learning, Reinforcement Learning, Machine Learning, Computer Vision, Natural Language Processing, Time Series Analytics, Data Mining, Data Analysis, Predictive Modelling. Work Experience Exafluence Inc., Data Scientist Intern February 2020 – May 2020 • Anomaly Detection System (Anomaly Detection, Python, KMeans, One-C SVM, GMM, R Shiny): - Designed and built Anomaly Detection ML Model using KMeans to achieve 0.9 F1 Score. - Implemented Deep Auto-Encoder Gaussian Mixture Model using TensorFlow from scratch (paper). - Designed and Developed a demo application in RShiny - Demo Rutgers University, Research Intern May 2019 – September 2019 • Adaptive Real Time Machine Learning Platform - ARTML (Python, CUDA, TensorFlow, CuPy, Numba, PyTorch): - Designed and Built GPU modules for ARTML using CuPy library, with 50% performance boost over CPU modules.
    [Show full text]
  • Tensorflow, Pytorch and Horovod
    Tensorflow, Pytorch and Horovod Corey Adams Assistant Computer Scientist, ALCF 1 Argonne Leadership1 ComputingArgonne Leadership Facility Computing Facility [email protected] Corey Adams Prasanna Taylor Childers Murali Emani Elise Jennings Xiao-Yong Jin Murat Keceli Balaprakash Adrian Pope Bethany Lusch Alvaro Vazquez Mayagoitia Misha Salim Himanshu Sharma Ganesh Sivaraman Tom Uram Antonio Villarreal Venkat Vishwanath Huihuo Zheng 2 Argonne Leadership Computing Facility ALCF Datascience Group Supports … • Software: optimized builds of important ML and DL software (tensorflow, pytorch, horovod) • Projects: datascience members work with ADSP projects, AESP projects, and other user projects to help users deploy their science on ALCF systems • Users: we are always interested to get feedback and help the users of big data and learning projects, whether you’re reporting a bug or telling us you got great performance. 3 Argonne Leadership Computing Facility What is Deep Learning? Photo from Analytics Vidhya https://thispersondoesnotexist.com Deep learning is … • an emerging exploding field of research that is transforming how we do science. • able to outperform humans on many complex tasks such as classification, segmentation, regression • able to replace data and simulation with hyper realistic generated data • expensive and time consuming to train 4 Argonne Leadership Computing Facility Deep Learning and Machine Learning on Aurora Aurora will be an Exascale system highly optimized for Deep Learning • Intel’s discrete GPUs (Xe) will drive accelerated single node performance • Powerful CPUs will keep your GPUs fed with data and your python script moving along quickly • High performance interconnect will let you scale your model training and inference to large scales. • Optimized IO systems will ensure you can keep a distributed training fed with data at scale.
    [Show full text]
  • Numba: a Compiler for Python Functions
    Numba: A Compiler for Python Functions Stan Seibert Director of Community Innovation @ Anaconda © 2018 Anaconda, Inc. My Background • 2008: Ph.D. on the Sudbury Neutrino Observatory • 2008-2013: Postdoc working on SNO, SNO+, LBNE (before it was DUNE), MiniCLEAN dark matter experiment • 2013-2014: Worked at vehicle tracking and routing company (multi-traveling salesman problem) • 2014-present: Employed at Anaconda (formerly Continuum Analytics), the creators of the Anaconda Python/R Distribution, as well many open source projects. Currently, I: • Manage the Numba project • Coordinate all of our open source projects • Bunch of other things (it's a startup!) © 2018 Anaconda, Inc. A Compiler for Python? Striking a Balance Between Productivity and Performance • Python has become a very popular language for scientific computing • Python integrates well with compiled, accelerated libraries (MKL, TensorFlow, ROOT, etc) • But what about custom algorithms and data processing tasks? • Our goal was to make a compiler that: • Works within the standard Python interpreter, and does not replace it • Integrates tightly with NumPy • Compatible with both multithreaded and distributed computing paradigms • Can be targeted at non-CPU hardware © 2018 Anaconda, Inc. Numba: A JIT Compiler for Python • An open-source, function-at-a-time compiler library for Python • Compiler toolbox for different targets and execution models: • single-threaded CPU, multi-threaded CPU, GPU • regular functions, “universal functions” (array functions), etc • Speedup: 2x (compared to basic NumPy code) to 200x (compared to pure Python) • Combine ease of writing Python with speeds approaching FORTRAN • BSD licensed (including GPU compiler) • Goal is to empower scientists who make tools for themselves and other scientists © 2018 Anaconda, Inc.
    [Show full text]
  • Or How to Avoid Recoding Your (Entire) Application in Another Language Before You Start
    Speeding up Python Or how to avoid recoding your (entire) application in another language Before you Start ● Most Python projects are “fast enough”. This presentation is not for them. ● Profle your code, to identify the slow parts AKA “hot spots”. ● The results can be surprising at times. ● Don’t bother speeding up the rest of your code. Improving your Python ● Look into using a better algorithm or datastructure next. ● EG: Sorting inside a loop is rarely a good idea, and might be better handled with a tree or SortedDict. Add Caching ● Memorize expensive-to-compute (and recompute) values. ● Can be done relatively easily using functools.lru_cache in 3.2+. Parallelise ● Multicore CPU’s have become common. ● Probably use threading or multiprocessing. Threading ● CPython does not thread CPU-bound tasks well. ● CPython can thread I/O-bound tasks well. ● CPython cannot thread a mix of CPU-bound and I/ O-bound well. ● Jython, IronPython and MicroPython can thread arbitrary workloads well. ● Data lives in the same threaded process. This is good for speed, bad for reliability. Multiprocessing ● Works with most Python implementations. ● Is a bit slower than threading when threading is working well. ● Gives looser coupling than threading, so is easier to “get right”. ● Can pass data from one process to another using shared memory. Numba ● If you’re able to isolate your performance issue to a small number of functions or methods (callables), consider using Numba. ● It’s just a decorator – simple. ● It has “object mode” (which doesn’t help much), and “nopython mode” (which can help a lot, but is more restrictive).
    [Show full text]