For some time now, we have noticed a high demand for gpu nodes, which increases the waiting time to launch your jobs.
We know how important your projects are, so here are some tips to improve the way you request resources :
Check the gpu partition on both cluster:
It often happens that the gpu of baobab are more solicited than those of yggdrasil.
Note that your fairshare is not shared on both clusters, so you will have different priorities depending on your usage on each cluster.
Evaluate your needs as best as possible:
To optimize the resources usage on the cluster, you need to know what you need for your job:
- The number of GPU.
- The number of CPU.
- The amount of memory and memory per GPU or CPU.
- The execution time.
- Use specific GPU depending your real needs to avoid underutilization
Example: Restrict yourself to low end GPU model
You also can use the debug-gpu partition on Yggdrasil to test your settings.
If you need advises, you can post on hpc-community or send us an e-mail.
New GPU coming soon:
In January, 9 GPU servers will be installed on Baobab for a total of 72 additional gpu cards.