[YARN-11452] Ability to use GPUs on YARN Nodemanager when running inside Docker

• Type: Bug
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 3.2.4
• Component/s: nodemanager

Environment

Docker

• Target Version/s: 3.2.4
• Assignee: Unassigned
• Reporter: Michael Griffiths
• Created: 07/Mar/23 03:50
• Updated: 07/Mar/23 03:50

Description

When running Nodemanagers in docker I have tried to enable GPU support however this fails because it seems to require GPU isolation to be enabled.  GPU isolation requires cgroups which are not available in Docker containers (I even tried using lxcfs to simulate cgroups - but this also fails with an error about mounting the /cgroups/devices/hadoop-yarn directory).

Is it possible to configure Nodemanagers so that they can detect/use GPUs without enabling GPU isolation?