GPU Computing for Data Science
John Joo
[email protected]
Data Science Evangelist @ Domino Data Lab
Outline
• Why use GPUs?
• Example applications in data science
• Programming your GPU
Case Study:
Monte Carlo Simulations
• Simulate behavior when randomness
is a key component
• Average the results of many
simulations
• Make predictions
Little Information in One “Noisy Simulation”
Price(t+1) = Price(t) e InterestRate•dt + noise
Many “Noisy Simulations” ➡ Actionable Information
Price(t+1) = Price(t) e InterestRate•dt + noise
Monte Carlo Simulations Are Often Slow
• Lots of simulation data is required to
create valid models
• Generating lots of data takes time
• CPU works sequentially
CPUs designed for sequential, complex tasks
Source: Mythbusters https://youtu.be/-P28LKWTzrI
GPUs designed for parallel, low level tasks
Source: Mythbusters https://youtu.be/-P28LKWTzrI
GPUs designed for parallel, low level tasks
Source: Mythbusters https://youtu.be/-P28LKWTzrI
Applications of GPU Computing in Data Science
• Matrix Manipulation
Algorithms for GPU Acceleration
• Numerical Analysis
• Inherently parallel
• Sorting
• Matrix operations
• FFT
• High FLoat-point Operations Per Sec
• String matching
(FLOPS)
• Monte Carlo simulations
• Machine learning
• Search
GPUs Make Deep Learning Accessible
Google
Stanford AI Lab
Datacenter
# of machines 1,000 3
# of CPUs or
2,000 CPUs 12 GPUs
GPUs
Cores 16,000 18,432
Power used 600 kW 4 kW
Cost $5,000,000 $33,000
Adam Coates, Brody Huval, Tao Wang, David Wu, Bryan Catanzaro, Ng Andrew ; JMLR W&CP 28 (3) : 1337–1345, 2013
CPU vs GPU Architecture:
Structured for Different Purposes
CPU GPU
4-8 High Performance Cores 100s-1000s of bare bones cores
Both CPU and GPU are required
Compute intensive
functions
Everything else
CPU GPU
General Purpose GPU Computing (GPGPU)
Heterogeneous Computing
Getting Started: Hardware
• Need a computer with GPU
• GPU should not be operating your
display
Spin up a GPU/CPU computer with 1 click.
8 CPU cores, 15 GB RAM
1,536 GPU cores, 4GB RAM
Getting Started: Hardware
✔
Getting Started: Software
Programming CPU Programming GPU
• Sequential • Parallel
• Write code top to bottom • Multi-threaded - race conditions
• Can do complex tasks • Low level tasks
• Independent • Dependent on CPU
Talking to your GPU
CUDA and OpenCL are GPU computing frameworks
Choosing How to Interface with GPU:
Simplicity vs Flexibility
High Application
specific
libraries
General
Simplicity purpose GPU
libraries
Custom
CUDA/
Low
OpenCL code
Low
Flexibility High
Application Specific Libraries
Python R
• Theano - Symbolic math • cudaBayesreg - fMRI
• TensorFlow - ML • mxnet - NN
• Lasagne - NN • rpud -SVM
• Pylearn2 - ML • rgpu - bioinformatics
• mxnet - NN
• ABSsysbio - Systems Bio Tutorial on using Theano, Lasagne, and no-learn:
http://blog.dominodatalab.com/gpu-computing-and-deep-learning/
General Purpose GPU Libraries
• Python and R wrappers for basic matrix
and linear algebra operations
• scikit-cuda
• cudamat
• gputools
• HiPLARM
• Drop-in library
Drop-in Library
Also works for Python!
http://scelementary.com/2015/04/09/nvidia-nvblas-in-numpy.html
Credit: NVIDIA
Custom CUDA/OpenCL Code
1. Allocate memory on the GPU
2. Transfer data from CPU to GPU
3. Launch the kernel to operate on the CPU
cores
4. Transfer results back to CPU
Example of using Python and CUDA:
Monte Carlo Simulations
• Using PyCuda to interface Python and
CUDA
• Simulating 3 million paths, 100 time steps
each
Python Code for CPU
Python/PyCUDA Code for GPU
8 more lines of code
Python Code for CPU
Python/PyCUDA Code for CPU
1. Allocate memory on the GPU
Python Code for CPU
Python/PyCUDA Code for CPU
2. Transfer data from CPU to GPU
Python Code for CPU
Python/PyCUDA Code for CPU
3. Launch the kernel to operate on the CPU cores
Python Code for CPU
Python/PyCUDA Code for CPU
4. Transfer results back to CPU
Python Code for CPU
26 sec
Python/PyCUDA Code for CPU
8 more lines of code
1.5 sec
17x speed up
Some sample Jupyter notebooks
• https://app.dominodatalab.com/johnjoo/gpu_examples
• Monte Carlo example using PyCUDA
• PyCUDA example compiling CUDA C for kernel
instructions
• Scikit-cuda example of matrix multiplication
• Calculating a distance matrix using rpud
More resources
• NVIDIA
• https://developer.nvidia.com/how-to-cuda-python
• Berkeley GPU workshop
• http://www.stat.berkeley.edu/scf/paciorek-
gpuWorkshop.html
• Duke Statistics on GPU (Python)
• http://people.duke.edu/~ccc14/sta-663/
CUDAPython.html
• Andreas Klockner’s webpage (Python)
• http://mathema.tician.de/
• Summary of GPU libraries
• http://fastml.com/running-things-on-a-gpu/
More resources
• Walk through of CUDA programming in R
• http://blog.revolutionanalytics.com/2015/01/parallel-
programming-with-gpus-and-r.html
• List of libraries for GPU computing in R
• https://cran.r-project.org/web/views/
HighPerformanceComputing.html
• Matrix computations in Machine Learning
• http://numml.kyb.tuebingen.mpg.de/numl09/
talk_dhillon.pdf
Questions?
[email protected] blog.dominodatalab.com
[email protected]
blog.dominodatalab.com
