[
https://issues.apache.org/jira/browse/MESOS-9691?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16809875#comment-16809875
]
Benjamin Mahler commented on MESOS-9691:
----------------------------------------
Re-opening as there is an issue with the fix.
> Quota headroom calculation is off when subroles are involved.
> -------------------------------------------------------------
>
> Key: MESOS-9691
> URL: https://issues.apache.org/jira/browse/MESOS-9691
> Project: Mesos
> Issue Type: Bug
> Components: allocation
> Reporter: Meng Zhu
> Assignee: Benjamin Mahler
> Priority: Critical
> Labels: resource-management
> Fix For: 1.8.0
>
>
> Quota "availableHeadroom" calculation:
> https://github.com/apache/mesos/blob/6276f7e73b0dbe7df49a7315cd1b83340d66f4ea/src/master/allocator/mesos/hierarchical.cpp#L1751-L1754
> is off when subroles are involved.
> Specifically, in the formula
> {noformat}
> available headroom = total resources - allocated resources - (total
> reservations - allocated reservations) - unallocated revocable resources
> {noformat}
> -The "allocated resources" part is hierarchical-aware and aggregate that
> across all roles, thus allocations to subroles will be counted multiple times
> (in the case of "a/b", once for "a" and once for "a/b").- Looks like due to
> the presence of `INTERNAL` node,
> `roleSorter->allocationScalarQuantities(role)` is *not* hierarchical. Thus
> this is not an issue.
> (If role `a/b` consumes 1cpu and `a` consumes 1cpu, if we query
> `roleSorter->allocationScalarQuantities("a");` It will return 1cpu, which is
> correct. In the sorter, there are four nodes, root, `a` (internal, 1cpu),
> `a/.` (leaf, 1cpu), `a/b` (leaf, 1cpu). Query `a` will return `a/.`)
> The "total reservations" is correct, since today it is "flat" (reservations
> made to "a/b" are not counted to "a"). Thus all reservations are only counted
> once -- which is the correct semantic here. However, once we fix MESOS-9688
> (which likely requires reservation tracking to be hierarchical-aware), we
> need to ensure that the accounting is still correct.
> -The "allocated reservations" is hierarchical-aware, thus overlap accounting
> would occur.- Similar to the `"allocated resources"` above, this is also not
> an issue at the moment.
> Basically, when calculating the available headroom, we need to ensure
> "single-counting". Ideally, we only need to look at the root's consumptions.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
