Using GPUs with Python

Questions

  • What is GPU acceleration?

  • How to enable GPUs (for instance with CUDA) in Python code?

  • How to deploy GPUs at HPC2N and UPPMAX?

Objectives

  • Learn common schemes for GPU code acceleration

  • Learn about the GPU nodes at HPC2N and UPPMAX

GPU-accelerated computing is when you use a graphics processing unit (GPU) along with a computer processing unit (CPU) to facilitate processing-intensive operations such as deep learning, analytics and engineering applications.

A GPU is a processor which is from many smaller and more specialized cores. When these cores work together, you can get a large performance boost for tasks that can be divided up and processed across many cores.

In a typical cluster, some GPUs are attached to a single node resulting in a CPU-GPU hybrid architecture. The CPU component is called the host and the GPU part the device.

We can characterize the CPU and GPU performance with two quantities: the latency and the throughput.

Latency refers to the time spent in a sole computation. Throughput denotes the number of computations that can be performed in parallel. Then, we can say that a CPU has low latency (able to do fast computations) but low throughput (only a few computations simultaneously). In the case of GPUs, the latency is high and the throughput is also high. We can visualize the behavior of the CPUs and GPUs with cars as in the figure below. A CPU would be compact road where only a few racing cars can drive whereas a GPU would be a broader road where plenty of slow cars can drive.

../_images/cpu-gpu-highway.png

Cars and roads analogy for the CPU and GPU behavior. The compact road is analogous to the CPU (low latency, low throughput) and the broader road is analogous to the GPU (high latency, high throughput)

Not every Python program is suitable for GPU acceleration. GPUs process simple functions rapidly, and are best suited for repetitive and highly-parallel computing tasks. GPUs were originally designed to render high-resolution images and video concurrently and fast, but since they can perform parallel operations on multiple sets of data, they are also often used for other, non-graphical tasks. Common uses are machine learning and scientific computation were the GPUs can take advantage of massive parallelism.

Many Python packages are not CUDA aware, but some have been written specifically with GPUs in mind. If you are usually working with for instance NumPy and SciPy, you could optimize your code for GPU computing by using CuPy which mimics most of the NumPy functions. Another option is using Numba, which has bindings to CUDA and lets you write CUDA kernels in Python yourself. This means you can use custom algorithms.

One of the most common use of GPUs with Python is for machine learning or deep learning. For these cases you would use something like Tensorflow or PyTorch libraries which can handle CPU and GPU processing internally without the programmer needing to do so.

GPUs on UPPMAX and HPC2N systems

There are generally either not GPUs on the login nodes or they cannot be accessed for computations. To use them you need to either launch an interactive job or submit a batch job.

UPPMAX only

Rackham’s compute nodes do not have GPUs. You need to use Snowy for that. A useful module on Snowy is python_ML_packages/3.9.5-gpu.

You need to use this batch command (for x being the number of cards, 1 or 2):

#SBATCH -M snowy
#SBATCH --gres=gpu:x

HPC2N

Kebnekaise’s GPU nodes are considered a separate resource, and the regular compute nodes do not have GPUs.

You need to use this to the batch system: #SBATCH --gres=gpu:<card>:x, for <card>=v100 or a100 and x=1 or 2.

And for the A100 GPUs you also need to use #SBATCH -p amd_gpu

Numba example

Numba is installed on HPC2N. We also need numpy, so we are loading SciPy-bundle as we have done before. We will use Python 3.9.x on both UPPMAX and HPC2N since that is the only version where the GPU packages are currently working correctly.

We are going to use the following program for testing (it was taken from https://linuxhint.com/gpu-programming-python/ but there are also many great examples at https://numba.readthedocs.io/en/stable/cuda/examples.html):

Python example add-list.py using Numba

import numpy as np
from timeit import default_timer as timer
from numba import vectorize

# This should be a substantially high value.
NUM_ELEMENTS = 100000000

# This is the CPU version.
def vector_add_cpu(a, b):
  c = np.zeros(NUM_ELEMENTS, dtype=np.float32)
  for i in range(NUM_ELEMENTS):
      c[i] = a[i] + b[i]
  return c

# This is the GPU version. Note the @vectorize decorator. This tells
# numba to turn this into a GPU vectorized function.
@vectorize(["float32(float32, float32)"], target='cuda')
def vector_add_gpu(a, b):
  return a + b;

def main():
  a_source = np.ones(NUM_ELEMENTS, dtype=np.float32)
  b_source = np.ones(NUM_ELEMENTS, dtype=np.float32)

  # Time the CPU function
  start = timer()
  vector_add_cpu(a_source, b_source)
  vector_add_cpu_time = timer() - start

  # Time the GPU function
  start = timer()
  vector_add_gpu(a_source, b_source)
  vector_add_gpu_time = timer() - start

  # Report times
  print("CPU function took %f seconds." % vector_add_cpu_time)
  print("GPU function took %f seconds." % vector_add_gpu_time)

  return 0

if __name__ == "__main__":
  main()

As before, we need the batch system to run the code. There are no GPUs on the login nodes.

Type-Along

Here we need to use the numba we installed in the “Example-gpu” virtual environment because of some temporary error (otherwise we would use the module python_ML_packages/3.9.5-gpu on Snowy) When you code-along, remember to change the activation path for the virtual environment to your own!

$ interactive -A naiss2024-22-107 -n 1 -M snowy --gres=gpu:1  -t 1:00:01 --gres=gpu:1  -t 1:00:01
You receive the high interactive priority.

Please, use no more than 8 GB of RAM.

Waiting for job 8483006 to start...
Starting job now -- you waited for 10 seconds.

$ ml python/3.9.5
$ source <path-to-virtual-environment>/Example-gpu/bin/activate
(Example-gpu) $ python add-list.py
CPU function took 36.849201 seconds.
GPU function took 1.574953 seconds.

Exercises

Integration 2D with Numba

An initial implementation of the 2D integration problem with the CUDA support for Numba could be as follows:

integration2d_gpu.py

from __future__ import division
from numba import cuda, float32
import numpy
import math
from time import perf_counter

# grid size
n = 100*1024
threadsPerBlock = 16
blocksPerGrid = int((n+threadsPerBlock-1)/threadsPerBlock)

# interval size (same for X and Y)
h = math.pi / float(n)

@cuda.jit
def dotprod(C):
    tid = cuda.threadIdx.x + cuda.blockIdx.x * cuda.blockDim.x

    if tid >= n:
        return

    #cummulative variable
    mysum = 0.0
    # fine-grain integration in the X axis
    x = h * (tid + 0.5)
    # regular integration in the Y axis
    for j in range(n):
        y = h * (j + 0.5)
        mysum += math.sin(x + y)

    C[tid] = mysum


# array for collecting partial sums on the device
C_global_mem = cuda.device_array((n),dtype=numpy.float32)

starttime = perf_counter()
dotprod[blocksPerGrid,threadsPerBlock](C_global_mem)
res = C_global_mem.copy_to_host()
integral = h**2 * sum(res)
endtime = perf_counter()

print("Integral value is %e, Error is %e" % (integral, abs(integral - 0.0)))
print("Time spent: %.2f sec" % (endtime-starttime))

Notice the larger size of the grid in the present case (100*1024) compared to the serial case’s size we used previously (10000). Large computations are necessary on the GPUs to get the benefits of this architecture.

One can take advantage of the shared memory in a thread block to write faster code. Here, we wrote the 2D integration example from the previous section where threads in a block write on a shared[] array. Then, this array is reduced (values added) and the output is collected in the array C. The entire code is here:

integration2d_gpu_shared.py

from __future__ import division
from numba import cuda, float32
import numpy
import math
from time import perf_counter

# grid size
n = 100*1024
threadsPerBlock = 16
blocksPerGrid = int((n+threadsPerBlock-1)/threadsPerBlock)

# interval size (same for X and Y)
h = math.pi / float(n)

@cuda.jit
def dotprod(C):
    # using the shared memory in the thread block
    shared = cuda.shared.array(shape=(threadsPerBlock), dtype=float32)

    tid = cuda.threadIdx.x + cuda.blockIdx.x * cuda.blockDim.x
    shrIndx = cuda.threadIdx.x

    if tid >= n:
        return

    #cummulative variable
    mysum = 0.0
    # fine-grain integration in the X axis
    x = h * (tid + 0.5)
    # regular integration in the Y axis
    for j in range(n):
        y = h * (j + 0.5)
        mysum += math.sin(x + y)

    shared[shrIndx] = mysum

    cuda.syncthreads()

    # reduction for the whole thread block
    s = 1
    while s < cuda.blockDim.x:
        if shrIndx % (2*s) == 0:
            shared[shrIndx] += shared[shrIndx + s]
        s *= 2
        cuda.syncthreads()
    # collecting the reduced value in the C array
    if shrIndx == 0:
        C[cuda.blockIdx.x] = shared[0]

# array for collecting partial sums on the device
C_global_mem = cuda.device_array((blocksPerGrid),dtype=numpy.float32)

starttime = perf_counter()
dotprod[blocksPerGrid,threadsPerBlock](C_global_mem)
res = C_global_mem.copy_to_host()
integral = h**2 * sum(res)
endtime = perf_counter()

print("Integral value is %e, Error is %e" % (integral, abs(integral - 0.0)))
print("Time spent: %.2f sec" % (endtime-starttime))

Prepare a batch script to run these two versions of the integration 2D with Numba support and monitor the timings for both cases.

Here follows a solution for HPC2N. Try and make it run on Snowy, by using a numba you install in a virtual environment and doing the changes suggested by the UPPMAX solution for add-list.py above.

Keypoints

  • You deploy GPU nodes via SLURM, either in interactive mode or batch

  • In Python the numba package is handy

Additional information