What Is the Impact of Resource Overcommitment on Notebook Instances?

Notebook overcommitment refers to the sharing of GPUs and memory within a node. To fully utilize resources, they are overcommitted in dedicated pools.

Example: A dedicated pool has one CPU node with 8 vCPUs and 64 GB memory. If you create a notebook instance with 2 vCPUs and 8 GB memory, a maximum of 6.67 notebook instances (8 vCPUs/(2 vCPUs x 0.6)) can be started due to overcommitment with an overcommitment ratio of 0.6. In this case, at least 1.2 vCPUs are required for starting the notebook instance, and a maximum of 2 vCPUs are used for running the notebook instance. Similarly, at least 4.8 GB memory is required, and a maximum of 8 GB memory is used for running the notebook instance.

Instances may be forcibly terminated due to overcommitment. For example, if six instances with 2 vCPUs are started on an 8 vCPUs node and the CPU usage of one instance exceeds the upper limit (8 vCPUs) of the node, Kubernetes forcibly terminates the instance that uses the most resources.

Do not overcommit resources as it may result in instance restart.