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, UPPMAX, LUNARC, NSC, and PDC?

Objectives

  • Get an intro to common schemes for GPU code acceleration

  • Learn about the GPU nodes at HPC2N, UPPMAX, LUNARC, NSC, PDC

  • Learn how to make a batch script asking for GPU nodes at HPC2N, UPPMAX, LUNARC, NSC, PDC

Introduction

In order to understand the capabilities of a GPU, it is instructive to compare a pure CPU architecture with a GPU based architecture. Here, there is a schemematics of the former:

../_images/AMD-Zen4-CPU-b-cn1701.png

Pure CPU architecture (single node). In the present case there are 256 cores, each with its own cache memory (LX). There is a shared memory (~378 GB/NUMA node) for all these cores. This is an AMD Zen4 node. The base frequency is 2.25 GHz, but it can boost up to 3.1 GHz.

As for the GPU architecture, a GPU card of type Ada Lovelace (like the L40s) looks like this:

../_images/lovelace-diagram.png

Note: The AD102 GPU also includes 288 FP64 Cores (2 per SM) which are not depicted in the above diagram. The FP64 TFLOP rate is 1/64th the TFLOP rate of FP32 operations. The small number of FP64 Cores are included to ensure any programs with FP64 code operate correctly, including FP64 Tensor Core code. This is a single GPU engine of a L40s card. There are 12 Graphics Processing Clusters (GPCs), 72 Texture Processing Clusters (TPCs), 144 Streaming Multiprocessors (SMs), and a 384-bit memory interface with 12 32-bit memory controllers). On the diagram, each green dot represents a CUDA core (single precision), while the yellow are RT cores and blue Tensor cores. The cores are arranged in the slots called SMs in the figure. Cores in the same SM share some local and fast cache memory.

../_images/GPC-with-raster-engine.png

The GPC is the dominant high-level hardware block. Each GPC includes a dedicated Raster Engine, two Raster Operations (ROPs) partitions, with each partition containing eight individual ROP units, and six TPCs. Each TPC includes one PolyMorph Engine and two SMs. Each SM contain 128 CUDA Cores, one Ada Third-Generation RT Core, four Ada Fourth-Generation Tensor Cores, four Texture Units, a 256 KB Register File, and 128 KB of L1/Shared Memory, which can be configured for different memory sizes depending on the needs of the graphics or compute workload.

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. One possible layout (Kebnekaise, AMD Zen4 node with L40s GPU) is as follows:

../_images/AMD-Zen4-GPU-1605.png

Schematics of a hybrid CPU-GPU architecture. A GPU L40s card is attached to a NUMA island which in turn contains 24 cores (AMD Zen4 CPU node with 48 cores total). The NUMA island and the GPUs are connected through a PCI-E interconnect which makes the data transfer between both components rather slow.

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

  • 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. We will talk more about that later in the course.

GPUs on UPPMAX, HPC2N, LUNARC, NSC, and PDC 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

Rackham’s compute nodes do not have GPUs. You need to use Snowy for that. A useful module on Snowy is python_ML_packages/3.11.8-gpu.

Snowy has Nvidia T4’s.

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

#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.

Kebnekaise has a great many different types of GPUs:

  • V100 (2 cards/node)

  • A40 (8 cards/node)

  • A6000 (2 cards/node)

  • L40s (2 or 6 cards/node)

  • A100 (2 cards/node)

  • H100 (4 cards/node)

  • MI100 (2 cards/node)

To access them, you need to use this to the batch system:

#SBATCH --gpus=x

where x is the number of GPU cards you want. Above are given how many are on each type, so you can ask for up to that number.

In addition, you need to add this to the batch system:

#SBATCH -C <type>

where type is

  • v100

  • a40

  • a6000

  • l40s

  • a100

  • h100

  • mi100

For more information, see HPC2N’s guide to the different parts of the batch system: https://docs.hpc2n.umu.se/documentation/batchsystem/resources/

LUNARC

LUNARC has Nvidia A100 GPUs and Nvidia A40 GPUs, but the latter ones are reserved for interactive graphics work on the on-demand system, and Slurm jobs should not be submitted to them.

Thus in order to use the A100 GPUs on Cosmos, add this to your batch script:

A100 GPUs on AMD nodes:

#SBATCH -p gpua100
#SBATCH --gres=gpu:1

These nodes are configured as exclusive access and will not be shared between users. User projects will be charged for the entire node (48 cores). A job on a node will also have access to all memory on the node.

A100 GPUs on Intel nodes:

#SBATCH -p gpua100i
#SBATCH --gres=gpu:<number>

where <number> is 1 or 2 (Two of the nodes have 1 GPU and two have 2 GPUs).

NSC

Tetralith has Nvidia T4 GPUs. In order to access them, add this to your batch script or interactive job:

#SBATCH -n 1
#SBATCH -c 32
#SBATCH --gpus-per-task=1

PDC

Dardel has 4 AMD Instinct™ MI250X á 2 GCDs per node.

You need to add this to your batch script or interactive job in order to access them:

#SBATCH -N 1
#SBATCH --ntasks-per-node=1
#SBATCH -p gpu

Numba example

Numba is installed on some of the centers as a module (HPC2N and LUNARC), on UPPMAX in python_ML_packages-gpu, but not on NSC except in a very old version. because of this we will use the virtual environment created earlier today at NSC.

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

As before, we need a batch script to run the code. There are no GPUs on the login node.

Note Type along!

Running a GPU Python code interactively - on Snowy.

$ interactive -A uppmax2025-2-296 -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.

salloc: Pending job allocation 9697978
salloc: job 9697978 queued and waiting for resources
salloc: job 9697978 has been allocated resources
salloc: Granted job allocation 9697978
salloc: Waiting for resource configuration
salloc: Nodes s195 are ready for job
 _   _ ____  ____  __  __    _    __  __
| | | |  _ \|  _ \|  \/  |  / \   \ \/ /   | System:    s195
| | | | |_) | |_) | |\/| | / _ \   \  /    | User:      bbrydsoe
| |_| |  __/|  __/| |  | |/ ___ \  /  \    |
 \___/|_|   |_|   |_|  |_/_/   \_\/_/\_\   |
 ###############################################################################

        User Guides: https://docs.uppmax.uu.se/

        Write to support@uppmax.uu.se, if you have questions or comments.

[bbrydsoe@s195 python]$ ml uppmax python/3.11.8 python_ML_packages/3.11.8-gpu
[bbrydsoe@s195 python]$ python add-list.py
CPU function took 35.272032 seconds.
GPU function took 1.324215 seconds.

Exercises

Integration 2D with Numba

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

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:

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

Keypoints

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

  • In Python the numba package is handy

Important

  • Of course, interactive mode could also be from inside Jupyter, VScode, spyder …

  • We will use GPUs more in the ML/DL section!

Additional information