Using GPUs¶
The nxg and sbg nodes contain GPU cards that can provide huge acceleration for certain types of parallel computing tasks, via the CUDA and OpenCL frameworks.
Access to GPU nodes
Access to GPU nodes is available for free to QM Researchers. Please contact us if you would like to use these nodes so we can add you to the allowed user list and help you get started with your initial GPU job submission. Note that access to GPU nodes is not permitted for Undergraduate and MSc students, or external users.
Applications with GPU support¶
There is a
considerable number of
scientific and analytical applications with GPU support. While some have GPU
support out of the box, such as Matlab and
Ansys, others may require specific GPU-ready
builds. These may appear in the module avail list with a -gpu suffix. If
you require GPU support adding to a specific application, please
submit a request for a GPU build and provide some test
data.
Be aware that not every GPU-capable application will run faster on a GPU for your code. For example, CP2K only has a GPU port of the DBCSR sparse matrix library. If you are not using this library in your code then you will not experience a performance boost.
Submitting jobs to GPU nodes¶
To request a GPU the -l gpu=<count> option should be used in your job
submission, and the scheduler will automatically select a GPU node. Note that
requests are handled per node, so a request for 64 cores and 2 GPUs will result
in 4 GPUs across two nodes. Examples are shown below.
Selecting a specific GPU type¶
For compatibility, you may optionally require a specific GPU type. For example,
CUDA version 8 predates the V100 GPU, and is not supported, so
-l gpu_type=kepler would select nodes using the K80 GPU instead. Conversely,
nodes with the Volta V100 GPU may be selected with -l gpu_type=volta, and
Ampere A100 nodes may be selected with -l gpu_type=ampere.
GPU card allocation¶
Do not set the CUDA_VISIBLE_DEVICES variable
For reasons documented below, please do not set the CUDA_VISIBLE_DEVICES
variable in your job scripts.
Requesting cards with parallel PE
If using the parallel parallel environment requests will be exclusive,
please ensure that you correctly set slots and gpu to fill the node.
We have enabled GPU device cgroups (Linux Control Groups) across all GPU nodes on Apocrita, which means your job will only be presented the gpu cards which have been allocated by the scheduler, to prevent some applications from attaching to GPUs which have not been allocated to the job.
Previously, it was required to set the CUDA_VISIBLE_DEVICES variable in job
scripts to ensure the correct GPU is used in the job. However, this was a
workaround until the GPU device cgroups were applied. You should no longer
set this variable in your job scripts.
Inside your job, the GPU cards presented to your job will always appear as device 0 to device N, depending on how many GPU cards you have requested. Below demonstrates the devices presented to jobs, per GPU resources request:
| GPUs Requested | Devices Presented |
|---|---|
| 1 | 0 |
| 2 | 0, 1 |
| 3 | 0 - 2 |
| 4 | 0 - 3 |
Checking GPU usage¶
GPU usage can be checked with the nvidia-smi command e.g.:
$ nvidia-smi -l 1
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 Tesla V100-PCIE... On | 00000000:06:00.0 Off | 0 |
| N/A 70C P0 223W / 250W | 12921MiB / 16160MiB | 97% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 1 Tesla V100-PCIE... On | 00000000:2F:00.0 Off | 0 |
| N/A 30C P0 23W / 250W | 4MiB / 32510MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 2 Tesla V100-PCIE... On | 00000000:86:00.0 Off | 0 |
| N/A 30C P0 23W / 250W | 6MiB / 32510MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
| 3 Tesla V100-PCIE... On | 00000000:D8:00.0 Off | 0 |
| N/A 31C P0 23W / 250W | 6MiB / 32510MiB | 0% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 1557 C python 12915MiB |
+-----------------------------------------------------------------------------+
In this example we can see that the process is using GPU 0. We use the -l 1
option which tells nvidia-smi to repeatedly output the status. Should this
be run inside a job, GPU 0 would be the card you have been allocated, which
might not be system device 0.
If you SSH into a GPU node and run nvidia-smi, you will see all system
GPU devices by their real ID, rather than the allocated device number.
Similarly, the SGE_HGR_gpu environment variable inside jobs and the
qstat -j JOB_ID command will also show the actual GPU device granted.
Example job submissions¶
The following examples show the basic outline of job scripts for GPU nodes.
Note that while the general rule for compute nodes is to strictly request only
the cores and RAM you will be using, our GPU jobs are GPU-centric: request only
the GPUs you will be using, but select 8 cores per GPU, and 7.5GB per core. You
can increase this to 11GB if you are sure that your job will not use an nxg
node (e.g. by using -l gpu_type='volta|ampere'), as these have less RAM than
sbg nodes. More detailed examples can also be found on the application-specific
pages on this site (e.g.
TensorFlow)
h_vmem does not need to account for GPU RAM
The h_vmem request only refers to the system RAM, and does not need to
account for GPU RAM used. The full GPU RAM is automatically granted when
you request a GPU
Request one GPU¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 8 # 8 cores (32 per gpu node)
#$ -l h_rt=240:0:0 # 240 hours runtime
#$ -l h_vmem=7.5G # 7.5 * 8 = 60G
#$ -l gpu=1 # request 1 GPU per host
./run_code.sh
Request two GPUs on the same box¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 16 # 16 cores (32 per gpu node)
#$ -l h_rt=240:0:0 # 240 hours runtime
#$ -l h_vmem=7.5G # 7.5 * 16 = 120G
#$ -l gpu=2 # request 2 gpu per host
./run_code.sh
Request all GPUs on a node with 4 GPUs¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 32 # 32 cores (32 per gpu node)
#$ -l h_rt=240:0:0 # 240 hours runtime
#$ -l gpu=4 # request 4 gpu per host
#$ -l exclusive # request exclusive access to the node
./run_code.sh
If you are requesting all GPUs on a node, then choosing exclusive mode will give you access to all of the resources. Note that requesting a whole node will likely result in a long queueing time, unless you have access to an "owned" node that your research group has purchased.
Submitting jobs to an owned node¶
If your research group has purchased a node, the scheduler default action will be to check both the main GPU nodes and any owned GPU nodes for available slots. If you want to restrict your job to your owned nodes only (e.g. for performance, or to ensure consistency), then adding:
#$ -l owned
to the resource request section at the top of your job script will restrict the job to running on owned nodes only.
Getting help¶
If you are unsure about how to configure your GPU jobs, please contact us for assistance.