Running interactively on AccelerateAI

While we recommend using the batch queues for intensive workloads, we also have capacity for some interactive use of GPUs to test workflows and validate models before submitting larger jobs to the batch queues.

The A100 GPU is partitioned using Multi-Instance GPU; each interactive GPU has either 5GB or 10GB of memory, and 1/7 or 2/7 of the compute units of the full GPU respectively.

There are two routes to running interactively on AccelerateAI:

• JupyterHub: This allows you to write and run Jupyter Notebooks that have direct access to the GPU.
• Command-line: this allows you to run scripts and other software directly at the command prompt. Unlike a batch job, you do not lose access to the GPU after each process completes.

Both are time-limited, as all jobs on AccelerateAI are. If your training is at the point that it is taking longer than the time limit, then we recommend packaging your workload for the batch queue, whose more powerful GPUs should let it complete more quickly, and which will allow you to checkpoint and resume your job more easily.

JupyterHub

JupyterHub is currently in the process of being installed. Please check back later for instructions on how to use it.

For the time being, we recommend trying the command-line interface. If you need urgent access to a Jupyter Notebook interface backed by a GPU, then please get in touch with SA2C Support

Command-line

To allocate and start an interactive session on AccelerateAI, run:

$ srun --pty --account=scwXXXX --gres=gpu:1 --partition=accel_ai_mig /bin/bash

(Replace scwXXXX with your Supercomputing Wales project identifier.)

From here, you can load any needed modules and run software as if you were running directly at a command prompt on the machine.

Once your interactive session is concluded, make sure to end the session with the exit command, so that the GPU is freed to allow others to use it.