[
https://issues.apache.org/jira/browse/UIMA-4655?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14974205#comment-14974205
]
Burn Lewis commented on UIMA-4655:
----------------------------------
The current design is to pro-actively request an extra reservation if the
number of free slots is less than this reserve.count and also to not waste
resources by releasing a reservation if there are at least this many free slots
in the remaining reservation(s). If the reservations are small then it is more
likely that the number of active jobs could bounce back and forth across this
request/release boundary, but this should only cause preemption churn if the
system is full, when we'd have churn anyway when incoming jobs cause
re-balancing of resources. Rather than trying to make the OR predict the
future job load, we could put the JD nodes at the bottom of the RM's search
order so they are used only when all others are full. Could we also avoid
putting other non-preemptable work on these nodes?
> Investigate if JD scheduler releases its extra reservations too frequently
> ---------------------------------------------------------------------------
>
> Key: UIMA-4655
> URL: https://issues.apache.org/jira/browse/UIMA-4655
> Project: UIMA
> Issue Type: Improvement
> Components: DUCC
> Reporter: Burn Lewis
> Priority: Minor
> Fix For: 2.1.0-Ducc
>
>
> A released reservation may be assigned to a job and then a short time later
> removed and handed back to the JD ... could cause churn.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
