Using resource groups with GPU hosts

Use dedicated resource groups for CPU hosts and GPU hosts. You can either create two new resource groups or use your existing resource groups. When you use existing resource groups, ensure that you specify one for CPU scheduling and the other for GPU scheduling.

Your cluster must include hosts that have GPUs.

Follow these steps to create new resource groups. If you plan to use existing resource groups, proceed to enable Spark instance groups that use GPU resources for applications.

If all GPUs on the machine are in default mode, which means sharable, you can select Number of slots per host and specify a number, which is n times the number of GPUs on the host. The bigger the value is, the more applications share the GPU. If all GPUs on the machine are in exclusive mode, select Advanced formula and enter ngpus as the value; so the slots number on the host is the number of GPUs on the host.

To create resource groups for slot-based scheduling, go to Resources > Resource Planning (Slot) > Resource Groups. For more information, see Creating resource groups.
Create two resource groups: one with CPU hosts (for example, CPUrg); the other with GPU hosts (for example, GPUrg). Do not use the same name for the two resource groups even if they contain the same hosts.
Note:
- For the resource group with the GPU hosts, ensure that all the hosts in the resource group have a number value in the ngpus column; and that the number under the Total slots for this group column is equal to the total number of GPUs on all the host in the resource group.
- If your cluster has non-GPU hosts, make sure to select Static (List of Names) as the Resource Selection Method and select the hosts from the Member hosts list. By default, the ngpus column is hidden but can be added using the Preferences menu.
After creating the resource groups, update the resource plan to use the resource groups you created. If required, modify the default resource plan to change how resources are allocated among consumers. For example, you can change the default slot allocation policy from Stacked to Exclusive. For more information, see Creating or modifying resource plans.
To create multidimensional resource groups for multidimensional scheduling, go to Resources > Resource Planning (Multidimensional) > Resource Groups. For more information, see Flow to configure multidimensional scheduling.
As part of this workflow, you create multidimensional resource groups, associate the resources groups with a multidimensional resource plan, add a consumer to the resource plan, and update other settings.

Create a Spark instance group. See Creating a Spark instance group to use GPUs.