[
https://issues.apache.org/jira/browse/FLINK-27921?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17557990#comment-17557990
]
Anton Kalashnikov commented on FLINK-27921:
-------------------------------------------
As I see at least in asyncWait benchmark, it recreates the job every iteration
so the scheduling phase is present on each iteration as well. And the benchmark
includes time for scheduling to the total time of the benchmark. Perhaps, it is
not good for the benchmark since the ideal benchmark should measure the only
time of targeted operation rather than all lifecycle(setup, shutdown) but it is
what we have right now.
First of all, we should be sure that it is the problem only for our benchmark
but it won't be visible to the user. If so, perhaps we can change the benchmark
configuration in FlinkEnvironmentContext to disable this delay for all
benchmarks. It looks not so good since we will benchmark non-default
configuration but at the same time, changing the default configuration, only
because our benchmark is not well written, is not a good idea as well.
[~roman], [~pnowojski], Maybe you have idea(or know), of how to make deal with
situations when our benchmark takes into account too many operations instead of
targeted ones. for example, instead of measuring only the time for processing
events it also takes into account the time of initial phase of the job. As
result, if the initial phase has an expected delay(exactly as this ticket),
then benchmark shows the degradation but in fact, there is no degradation in
the targeted operation.
> Introduce the checkResourceRequirementsWithDelay in DeclarativeSlotManager
> --------------------------------------------------------------------------
>
> Key: FLINK-27921
> URL: https://issues.apache.org/jira/browse/FLINK-27921
> Project: Flink
> Issue Type: Sub-task
> Components: Runtime / Coordination
> Reporter: Aitozi
> Assignee: Aitozi
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.16.0
>
>
> As discussed in
> [https://github.com/apache/flink/pull/19840#discussion_r884242067] .This
> ticket is meant to introduce the same mechanism to wait for a slight delay
> before process the resource check with {{{}FineGrainedSlotManager{}}}. It
> will reduce the frequency of unnecessary re-allocations
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
