[YARN-8911] ContainerScheduler incorrectly uses percentage number as the cpu resource utlization

• Type: Bug
• Status: Patch Available
• Priority: Major
• Resolution: Unresolved
• Component/s: nodemanager
• Assignee: Haibo Chen
• Reporter: Haibo Chen
• Created: 18/Oct/18 20:43
• Updated: 24/Oct/18 14:59

Description

UPDATE:  per discussion below, the cpu resource utlization (ResourceUtilzation.cpu) is incorrectly documented as a percentage number in [0, 1.0] in the java doc. It is however, used as the number of vcores used throughout the system.  YARN-4597 is a victim of the  misleading java doc.

 

 

ResourceUtilization represents the cpu utilization with a float number in [0, 1.0], i.e. the percentage of cpu usage across the node.  However, when Containers Monitor tracks the total aggregate resource utilization of all containers, it adds up the total number of vcores used by all running containers.

 

(If you have 6 containers running, each consuming 1 vcore, you'd see the aggregated cpu container utilization being 6.0, but it's supposed to be always between 0 and 1.0)   See the code

Issue Links

• relates to: YARN-8932 ResourceUtilization cpu is misused in oversubscription as a percentage