Motivation  

• We need this to ensure atomicity of accelerator’s hardware perspective.

Required Features

• Add new field under scaling_group.scheduler_opts JSON column indicating whether to allow fragmentation of accelerators supporting fractional shares
  • This should be an opt-in feature
• Check the newly created field on agent-side accelerator scheduling process and cancel the kernel creation process if there isn’t enough resource to avoid resource fragmentation

Impact  

• Administrators will be able to select whether to allow fragmentation or not on a scaling group basis

Testing Scenarios  

• Create a Backend.AI cluster consisted of multiple mock accelerators
• Try to create a session with accelerator forced to be fragmentized
• Verify kernel creation being blocked when the new option is enabled