GPU-accelerated machine learning

This guide explains the basics of using GPUs in CSC's supercomputers. It is part of our Machine learning guide.

Puhti, Mahti or LUMI?

Puhti and Mahti are CSC's two national supercomputers. Puhti has the largest number of GPUs (V100) and offers the widest selection of installed software, while Mahti has a smaller number of faster newer generation A100 GPUs.

The CSC-hosted European supercomputer LUMI will provide a massive GPU resource once it's up and running. Currently, only a small early access platform is available for testing code compatibility.

The main GPU-related statistics are summarized in the table below.

	GPU type	GPU memory	GPU nodes	GPUs/node	Total GPUs
Puhti	NVIDIA Volta V100	32 GB	80	4	320
Mahti	NVIDIA Ampere A100	40 GB	24	4	96
LUMI	AMD MI250x	64 (128) GB	2560	8 (4)	20480 (10240)

Note

Each LUMI node has 4 MI250x GPUs, however 8 GPUs will be available through Slurm as the MI250x card features 2 GPU dies (GCDs). The table above shows the GPU die specific numbers, MI250x card specific numbers are shown in parenthesis.

Please read our usage policy for the GPU nodes. Also consider that the Slurm queuing situation may vary between Puhti and Mahti at different times. However, note that Puhti and Mahti have distinct file systems, so you need to manually copy your files if you wish to change the system. In case you are unsure which supercomputer to use, Puhti is a good default as it has a wider set of software supported.

Available machine learning software

We support a number of applications for GPU-accelerated machine learning on CSC's supercomputers, including TensorFlow and PyTorch. Please read the detailed instructions for the specific application that you are interested in.

You need to use the module system to load the application you want, for example:

module load tensorflow/2.7

Please note that our modules already include CUDA and cuDNN libraries, so there is no need to load cuda and cudnn modules separately!

Finally, on Puhti, we provide some special applications which are not shown by default in the module system. These have been made available due to user requests, but with limited support. You can enable them by running:

module use /appl/soft/ai/singularity/modulefiles/

Installing your own software

In many cases, our existing modules provide the required framework, but some packages are missing. In this case you can often load the appropriate module and then install additional packages for personal use with the pip package manager.

For more complex software requirements, we recommend using Singularity containers. These are similar to Docker containers, and are well supported in CSC's supercomputers, in fact many of our own modules are based on Singularity. See our documentation on how to run Singularity containers, and how to create your own containers.

Running GPU jobs

To submit a GPU job to the Slurm workload manager, you need to use the gpu partition on Puhti or gpusmall or gpumedium on Mahti, and specify the type and number of GPUs required using the --gres flag. Below are example batch scripts for reserving one GPU and a corresponding 1/4 of the CPU cores of a single node:

PuhtiMahti

#!/bin/bash
#SBATCH --account=<project>
#SBATCH --partition=gpu
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=10
#SBATCH --mem=64G
#SBATCH --time=1:00:00
#SBATCH --gres=gpu:v100:1

srun python3 myprog.py <options>

#!/bin/bash
#SBATCH --account=<project>
#SBATCH --partition=gpusmall
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --time=1:00:00
#SBATCH --gres=gpu:a100:1

srun python3 myprog.py <options>

Mahti's gpusmall partition supports only jobs with 1-2 GPUs. If you need more GPUs, use the gpumedium queue. You can read more about multi-GPU and multi-node jobs in our separate tutorial.

For more detailed information about the different partitions, see our page about the available batch job partitions on CSC's supercomputers.

GPU utilization

GPUs are an expensive resource compared to CPUs (60 times more BUs!). Hence, ideally, a GPU should be maximally utilized once it has been reserved.

You can check the current GPU utilization of a running job by sshing to the node where it is running and running nvidia-smi. You should be able to identify your run from the process name or the process id, and check the corresponding "GPU-Util" column. Ideally it should be above 90%.

For example, from the following excerpts we can see that on GPU 0 a python3 job is running which uses roughly 17 GB of GPU memory and the current GPU utilization is 99% (i.e., very good).

| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla V100-SXM2...  On   | 00000000:61:00.0 Off |                    0 |
| N/A   51C    P0   247W / 300W |  17314MiB / 32510MiB |     99%      Default |

| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0    122956      C   python3                                    17301MiB |

Alternatively, you can use seff which shows GPU utilization statistics for the whole running time.

seff <job_id>

In this example we can see that maximum utilization is 100%, but average is 92%.

GPU load 
     Hostname        GPU Id      Mean (%)    stdDev (%)       Max (%) 
       r01g07             0         92.18         19.48           100 
------------------------------------------------------------------------
GPU memory 
     Hostname        GPU Id    Mean (GiB)  stdDev (GiB)     Max (GiB) 
       r01g07             0         16.72          1.74         16.91 

As always, don't hesitate to contact our service desk if you need advice on how to improve you GPU utilization.

Using multiple CPUs for data pre-processing

One common reason for the GPU utilization being low is when the CPU cannot load and pre-process the data fast enough, and the GPU has to wait for the next batch to process. It is then a common practice to reserve more CPUs to perform data loading and pre-processing in several parallel threads or processes. A good rule of thumb in Puhti is to reserve 10 CPUs per GPU (as there are 4 GPUs and 40 CPUs on each node). On Mahti you can reserve up to 32 cores, as that corresponds to 1/4 of the node. Remember that CPUs are a much cheaper resource than the GPU!

You might have noticed that we have already followed this advice in our example job scripts:

#SBATCH --cpus-per-task=10

Your code also has to support parallel pre-processing. However, most high-level machine learning frameworks support this out of the box. For example in TensorFlow you can use tf.data and set num_parallel_calls to the number of CPUs reserved and utilize prefetch:

dataset = dataset.map(..., num_parallel_calls=10)
dataset = dataset.prefetch(buffer_size)

In PyTorch, you can use torch.utils.DataLoader, which supports data loading with multiple processes:

train_loader = torch.utils.data.DataLoader(..., num_workers=10)

If you are using multiple data loaders, but data loading is still slow, it is also possible that you are using the shared file system inefficiently. A common error is to read a huge number of small files. You can read more about how to store and load data in the most efficient way for machine learning in our separate tutorial.

Profilers

TensorFlow Profiler and PyTorch Profiler are available as TensorBoard plugins. The profilers can be found at the PROFILE and PYTORCH_PROFILER tabs in TensorBoard, respectively. Note that the tabs may not be visible by default but can be found at the pull-down menu on the right-hand side of the interface. The profilers can be used to identify resource consumption and to resolve performance bottlenecks, in particular the data input pipeline.

See also how to launch TensorBoard using the Puhti web interface. The TensorFlow module tensorflow/2.8 or later is required to use the profilers.

---

Last update: October 7, 2022