GPU Computing with Apache Spark and Python

Stan Seibert Siu Kwan Lam

April 5, 2016

• 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 of Global State

• Beware that PySpark may (quite often!) spawn new processes

• New processes may have the same initial random seed

• the python random module must be seeded properly in each newly spawned process

• Use memoize

• example: use global dictionary to track shared states and remember performed initializations

51 Spark and Multi-GPU

• Spark is not aware of GPU resources

• Efficient usage of GPU resources require some workarounds e.g.

• option 1: random GPU assignment numba.cuda.select_device(random.randrange(num_of_gpus))

• option 2: use mapPartitionWithIndex selected_gpu = partition_index % num_of_gpus

• option 3: manage GPUs externally; like CaffeOnSpark

52 GPU assignment by partition index

Delegate GPU work to external GPU-aware process >2x speedup

54 PySpark and Numba for GPU clusters

• Numba let’s you create compiled CPU and CUDA functions right inside your Python applications.

• Numba can be used with Spark to easily distribute and run your code on Spark workers with GPUs

• There is room for improvement in how Spark interacts with the GPU, but things do work.

• Beware of accidentally multiplying fixed initialization and compilation costs.