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.
Short queue GPU testing
QM Researchers may consider testing their GPU code before requesting GPU access. We provide a short queue testing node "dn150" which supports GPU jobs up to 1 hour - see below for more information and examples.
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. The scheduler will automatically select a GPU node based on
availability and other resource parameters, if specified.
We also provide a node (dn150) with 2 Kepler K40c GPU cards which may be used before acquiring access to run on the other GPU nodes. This node has been designed to avoid long queuing times during GPU code testing and therefore, only accepts jobs submitted with up to 1 hour runtime, on the short queue.
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.
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
Testing GPU Node (dn150)¶
Memory requests during testing
The GPU testing node "dn150" only has a small amount of RAM. Therefore, jobs must request a maximum of 1G per core otherwise they will not be eligible to run on dn150. Jobs without the correct memory setting may be subject to increased job queuing times and/or produce undesirable results.
Request one dn150 GPU¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 8 # 8 cores (16 available on dn150)
#$ -l h_rt=1:0:0 # 1 hour runtime (required to run on dn150)
#$ -l h_vmem=1G # see above notice
#$ -l gpu=1 # request 1 GPU
./run_code.sh
Request two dn150 GPUs (whole node)¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 16 # 16 cores (16 available on dn150)
#$ -l h_rt=1:0:0 # 1 hour runtime (required to run on dn150)
#$ -l gpu=2 # request 2 GPUs
#$ -l exclusive # request exclusive access
./run_code.sh
Production GPU Nodes¶
Request one GPU¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 8 # 8 cores (32 available 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
./run_code.sh
Request two GPUs¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 16 # 16 cores (32 available 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 GPUs
./run_code.sh
Request four GPUs (whole node)¶
#!/bin/bash
#$ -cwd
#$ -j y
#$ -pe smp 32 # 32 cores (32 available per GPU node)
#$ -l h_rt=240:0:0 # 240 hours runtime
#$ -l gpu=4 # request 4 GPUs
#$ -l exclusive # request exclusive access
./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 (except dn150), unless you have access to an "owned" GPU node that your research group has purchased.
Submitting jobs to an owned node¶
If your research group has purchased a GPU node, the scheduler default action will be to firstly check for available slots on owned node(s), and then the public GPU nodes (if applicable). 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.