GPU computing

Introduction

GPUs, short for graphical processing unit, are massively-parallel processors that are optimized to perform parallel operations. Computations that might take days to run on CPUs, take substantially less time on GPUs. This speed-up specially comes in handy when dealing with large amounts of data, e.g. in machine learning/deep learning tasks, which is why GPUs have become an indispensable tool in the research community.

The programs we normally write in common programming languages, e.g. C++ are executed by the CPU. We need to explicitly communicate with the GPU if we want GPU to execute the program. That is, upload the program and the input data to the GPU, and transfer the result from the GPU to the main memory. What enable this procedure are programming environments designed to communicate with GPUs in such a manner. An example of such an API is CUDA which is the native programming interface for NVIDIA GPUs.

On Narvi, we have a large number of NVIDIA GPU cards from different generations and currently only support CUDA. Triton GPUs are not the typical desktop GPUs, but specialized research-grade server GPUs with large memory, high bandwidth and specialized instructions, that are constantly increasing in number. For scientific purposes, they generally outperform the best desktop GPUs.

See also

Please ensure you have read interactive and serial before you proceed with this tutorial.

Access

Currently all GPU-nodes are on gpu-partition and access to resources is limited to GPU usergroup. So if you are willing to use GPU’s, you’ll have to request access to GPU-group via email to tcsc.tau@tuni.fi

GPU jobs

To request GPUs on Slurm, you should use the --gres option either in your batch script or as a command-line argument to your interactive job. Used with a SBATCH directive in a batch script, exactly one GPU is requested as follows.

#SBATCH --gres=gpu:1 --partition=gpu

Note

This is narvi-specifics, on Aalto’s Triton you don’t need to specify partition

You can request as many GPUs as you’d like using #SBATCH --gres=gpu:<n> wherein n denotes the number of the requested GPUs.

Note

Most of the time, using more than one GPU isn’t worth it, unless you specially optimize, because communication takes too much time. It’s better to parallelize by splitting tasks into different jobs.

You can restrict yourself to a certain type of GPU card by using using the --constraint option. For example, to restrict to Kepler generation (K80s), use --constraint='kepler' or only Pascal or Volta generations with --constraint='pascal|volta' (Remember to use the quotes since | is the shell pipe)

Current GPU resources and architecture’s

Currently there are four different types of gpu’s available on Narvi: Tesla K40, K80, P100 and V100. Also there are some variation in memory amounts. If you wan’t to request specific type of GPU, here are suitable slurm –gres and –constraint -parameters:

GPU-type	mem/GPU	quantity	gres	constraint
Tesla K40	12GiB	6	teslak40	kepler
Tesla K80	12GiB	3	teslak80	kepler
Tesla P100	12GiB	24	teslap100	pascal
Tesla P100	16GiB	8	teslap100	gpumem_16,pascal
Tesla V100	16GiB	16	teslav100	gpumem_16, volta
Tesla V100	32GiB	24	teslav100	gpumem_32, volta
Quadro RTX 8000	4GB	3	rtx100	gpumem_44, rtx

So assuming you’ll want P100 with 16G, you should use something like –gres:gpu:teslap100:1 –constraint=gpumem_16 On the other hand, if anything after p100 is enough then you could ask –gres:gpu:1 –constraint=”pascal|volta”

Available machine learning frameworks

We support the following machine learning frameworks out of the box:

• Tensorflow: module load anaconda/2020-02-tf2. See the Tensorflow page for info on older versions.
• Keras: module load anaconda/2020-02-tf2
• PyTorch:module load anaconda/2020-02-tf2
• Detectron: via singularity images
• CNTK: via singularity images

Please note that most of the pre-installed softwares have CUDA already present. Thus you do not need to load CUDA as a seperate module when loading these. See the application list or GPU computing reference <https://scicomp.aalto.fi/triton/usage/gpu/> for more details.

Compiling CUDA-based code

To compile CUDA-based code for GPUs, you need to load the relevant cuda module. You can see what versions of CUDA is available using module spider:

module spider cuda

When submitting a batch script, you need to load the cuda module, compile your code, and subsequently run the executable. An example of such a submission script is shown below wherein the output of the code is written to a file named helloworld.out in the current directory:

#!/bin/bash
#SBATCH --time=00:05:00
#SBATCH --job-name=helloworld
#SBATCH --mem-per-cpu=500M
#SBATCH --cpus-per-task=1
#SBATCH --gres=gpu:1
#SBATCH --partition=gpu --reservation=FGCI2
#SBATCH --output=helloworld.out

module load cuda
nvcc helloworld.cu -o helloworld
./helloworld

Note

If you ever get libcuda.so.1: cannot open shared object file: No such file or directory, this means you are attempting to use a CUDA program on a node without a GPU. This especially happens if you try to test GPU code on the login node, and happens (for example) even if you try to import the GPU tensorflow module in Python on the login node.

Examples

Monitoring efficient use of GPUs

When running a GPU job, you should check that the GPU is being fully utilized. Additionally, for the sake of troubleshooting and ensuring that GPU is executing your code, not GPUs, you can run an interactive job:

srun --pty --gres=gpu:1 --partition=gpu --time=1:00:00 --mem=1G -c 3 /bin/bash

When assigned a node in the GPU partition, you can ssh to the node and run nvidia-smi. You can find your process by e.g. using htop and inspect the GPU-Util column. It should be close to 100%.

Once the job has finished, you can use slurm history to obtain the jobID and run:

sacct -j <jobID> -o comment -p

This also shows the GPU utilization. Also seff-command is patched to include GPU utilization off job:

seff -j <jobID>

Note

There are factors to be considered regarding efficient use of GPUs. For instance, is your code itself efficient enough? Are you using the framework pipelines in the intended fashion? Is it only using GPU for a small portion of the entire task? Amdahl’s law of parallelization speedup is relevant here.

Monitoring CPUs

If the GPU utilization of your jobs are low, you can do the seff <jobID> command and see if the CPU utilization is 100%. This could mean that the GPUs are not able to supply data fast enough.

Please keep in mind that when using a GPU, you need to also request enough CPUs to supply the data to the process. So, you can increase the number of CPUs you request so that enough data is provided for the GPU. However, you shouldn’t request too many: There wouldn’t be enough CPUs for everyone to use the GPUs, and they would go to waste (For the K80 nodes, we have only 1.5 CPUs per GPU, but on all others we have 4-6 CPUs/GPU).

Input/output

Deep learning work is intrinsically very data-hungry. Remember what we said about storage and input/output being important before (Data storage)? This matter becomes very important when working with GPUs. In fact, faster memory bandwidth is the main improvement of our server-grade GPUs compared to desktop models.

If you are loading big amounts of data, you should package the data into a container format first; lots of small files are your worst enemy. Each framework has a way to do this efficiently in a whole pipeline.

See also

Please refer to the small files <https://scicomp.aalto.fi/triton/usage/smallfiles/> page for more detailed information.

If your data consists of individual files that are not too big, it is a good idea to have the data stored in one file, which is then copied to nodes ramdisk /dev/shm or temporary disk /tmp.

If your data is too big to fit in the disk, we recommend that you contact us for efficient data handling models.

Exercises

1. Run nvidia-smi on a GPU node with srun. Use slurm history to check which GPU node you ended up on. Try setting a constraint to force a different GPU architecture.

2. Use hpc-examples/gpu/pi.cu from the previous exercises. Compile it using cuda module and nvcc. Run it. Does it say zero? Try running it with a GPU and see what happens.

3. Run one of the samples given above. Try using sbatch as well.

4. Modify CTNK sample slurm script in a way that it copies datasets to an unique folder in /dev/shm or $TMPDIR before running the Python code. Modify CNTK sample so that it loads data from the new location.

   HINT: Check out mktemp --help, command output substitutions-section from our Linux shell tutorial and the API page for Python’s os.environ.

   Solution to ex. 4: cntk_mnist_ex4.py cntk_mnist_ex4.sh.

What’s next?

You can see the main article, GPU Computing <https://scicomp.aalto.fi/triton/usage/gpu/>, for more detailed information regarding GPU computing, including examples of different machine learning frameworks.

This guide assumes you are using pre-existing GPU programs. If you need to write your own, that’s a whole other story, and you can find some hints on the reference page.