[SPARK-27488] Driver interface to support GPU resources

• Type: Story
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.0.0
• Fix Version/s: 3.0.0
• Component/s: Spark Core
• Epic Link: GPU-aware Scheduling
• Assignee: Thomas Graves
• Reporter: Thomas Graves
• Created: 17/Apr/19 14:08
• Updated: 23/May/19 18:50
• Resolved: 23/May/19 18:46

Description

We want to have an interface to allow the users on the driver to get what resources are allocated to them.  This is mostly to handle the case the cluster manager does not launch the driver in an isolated environment and where users could be sharing hosts.  For instance in standalone mode it doesn't support container isolation so a host may have 4 gpu's, but only 2 of them could be assigned to the driver.  In this case we need an interface for the cluster manager to specify what gpu's for the driver to use and an interface for the user to get the resource information

Issue Links

• blocks: SPARK-27823 Add an abstraction layer for accelerator resource handling to avoid manipulating raw confs
• is blocked by: SPARK-27024 Executor interface for cluster managers to support GPU resources
• is related to: SPARK-27364 User-facing APIs for GPU-aware scheduling
• links to: GitHub Pull Request #24615