wuyi created SPARK-26439:
----------------------------
Summary: Introduce WorkOffer reservation mechanism for Barrier
TaskSet
Key: SPARK-26439
URL: https://issues.apache.org/jira/browse/SPARK-26439
Project: Spark
Issue Type: Improvement
Components: Spark Core
Affects Versions: 2.4.0
Reporter: wuyi
Fix For: 2.4.0
Currently, Barrier TaskSet has a hard requirement that tasks can only be
launched
in a single resourceOffers round with enough slots(or sufficient resources), but
can not be guaranteed even if with enough slots due to task locality delay
scheduling.
So, it is very likely that Barrier TaskSet gets a chunk of sufficient resources
after
all the trouble, but let it go easily just beacuae one of pending tasks can not
be
scheduled. Futhermore, it causes severe resource competition between TaskSets
and jobs
and introduce unclear semantic for DynamicAllocation.
This pr trys to introduce WorkOffer reservation mechanism for Barrier TaskSet,
which
allows Barrier TaskSet to reserve WorkOffer in each resourceOffers round, and
launch
tasks at the same time once it accumulate the sufficient resource. In this way,
we
relax the requirement of resources for the Barrier TaskSet. To avoid the
deadlock which
may be introuduced by serveral Barrier TaskSets holding the reserved WorkOffer
for a
long time, we'll ask Barrier TaskSets to force releasing part of reserved
WorkOffers
on demand. So, it is highly possible that each Barrier TaskSet would be
launched in the
end.
To integrate with DynamicAllocation
The possible effective way I can imagine is that adding new event, e.g.
ExecutorReservedEvent, ExecutorReleasedEvent, which behaved like busy executor
with
running tasks or idle executor without running tasks. Thus,
ExecutionAllocationManager
would not let the executor go if it reminds of there're some reserved resource
on that
executor.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
