[YARN-10796] Capacity Scheduler: dynamic queue cannot scale out properly if its capacity is 0%

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 3.4.0
• Component/s: capacity scheduler, capacityscheduler
• Hadoop Flags: Reviewed
• Assignee: Peter Bacsko
• Reporter: Peter Bacsko
• Created: 31/May/21 18:28
• Updated: 03/Jun/21 11:28
• Resolved: 03/Jun/21 11:28

Description

If we have a dynamic queue (AutoCreatedLeafQueue) with capacity = 0%, then it cannot properly scale even if it's max-capacity and the parent's max-capacity would allow it.

Example:

Cluster Capacity:  16 GB / 16cpu (2 nodes, each with 8 GB / 8 cpu )
Container allocation size: 1G / 1 vcore

root.dynamic 
    Effective Capacity:      <memory: 8192, vCores: 8> ( 50.0%)
    Effective Max Capacity:  <memory:16384, vCores:16> (100.0%) 

    Template:
        Capacity:               40%
        Max Capacity:           100%
        User Limit Factor:      4
 

leaf-queue-template.capacity = 40%
leaf-queue-template.maximum-capacity = 100%
leaf-queue-template.maximum-am-resource-percent = 50%
leaf-queue-template.minimum-user-limit-percent =100%
leaf-queue-template.user-limit-factor = 4

"root.dynamic" has a maximum capacity of 100% and a capacity of 50%.

Let's assume there are running containers in these dynamic queues (MR sleep jobs):
root.dynamic.user1 = 1 AM + 3 container (capacity = 40%)
root.dynamic.user2 = 1 AM + 3 container (capacity = 40%)
root.dynamic.user3 = 1 AM + 15 container (capacity = 0%)

This scenario will result in an underutilized cluster. There will be approx 18% unused capacity. On the other hand, it's still possible to submit a new application to root.dynamic.user1 or root.dynamic.user2 and reaching a 100% utilization is possible.