[
https://issues.apache.org/jira/browse/FLINK-37412?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Matthias Pohl resolved FLINK-37412.
-----------------------------------
Release Note: A new strategy MIN_RESOURCES for selecting a TaskManager for
slot allocation on the ResourceManager side was introduced. This strategy
allows for optimizing the resource utilization, i.e. trying to allocate all
slots in one TaskManager before selecting another one (essentially, the
opposite of the fan-out strategy that can be configured via NODE)
Resolution: Fixed
master:
[034d19c63857c34406e0753a1ce522f0f5cab0bb|https://github.com/apache/flink/commit/034d19c63857c34406e0753a1ce522f0f5cab0bb]
> Add slot allocation strategy to for selecting TMs first that have already
> slots assigned
> ----------------------------------------------------------------------------------------
>
> Key: FLINK-37412
> URL: https://issues.apache.org/jira/browse/FLINK-37412
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Coordination
> Reporter: Matthias Pohl
> Assignee: Ahmed Hamdy
> Priority: Major
> Labels: pull-request-available
>
> Right now, the {{DefaultResourceAllocationStrategy}} has two different
> strategies for selecting the TM that should try to allocate a slot for a
> given Job:
> * {{LeastUtilizationResourceMatchingStrategy}}: Selects the TM based on it's
> utilization. TMs that have less resources allocated get precedence over more
> utilized TMs. This works as a fan-out strategy to spread the load onto as
> many TMs as possible.
> * {{AnyMatchingResourceMatchingStrategy}}: Kind of randomly selects the TM
> order to iterate over for allocating the resource requirements. This works
> kind of as a fallback/default.
> This Jira issue is about a third option: Trying to fully utilize TMs before
> assigning slots to new "untouched" TMs. This would help utilizing the
> available TMs.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
