[
https://issues.apache.org/jira/browse/FLINK-7219?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16093490#comment-16093490
]
Stephan Ewen commented on FLINK-7219:
-------------------------------------
[~sihuazhou] I think this one needs a bit more thought.
The future direction plane are more and more about state locality, meaning that
tasks go (on execution and on recovery) where their state was and is. Only when
there is no state constraint, then the inputs are picked.
> Current allocate strategy cann‘t achieve the optimal effect with input's
> location
> ---------------------------------------------------------------------------------
>
> Key: FLINK-7219
> URL: https://issues.apache.org/jira/browse/FLINK-7219
> Project: Flink
> Issue Type: Bug
> Components: JobManager
> Affects Versions: 1.3.1
> Reporter: Sihua Zhou
> Assignee: Sihua Zhou
>
> This is second subtask of issue
> [FLINK-7153|https://issues.apache.org/jira/browse/FLINK-7153?filter=-2].
> Current allocate strategy can't allocate the slot optimize. Here is the test
> case:
> {code}
> JobVertex v1 = new JobVertex("v1", jid1);
> JobVertex v2 = new JobVertex("v2", jid2);
> SlotSharingGroup group = new SlotSharingGroup();
> v1.setSlotSharingGroup(group);
> v2.setSlotSharingGroup(group);
> v1.setParallelism(2);
> v2.setParallelism(4);
> v1.setInvokableClass(BatchTask.class);
> v2.setInvokableClass(BatchTask.class);
> v2.connectNewDataSetAsInput(v1, DistributionPattern.POINTWISE,
> ResultPartitionType.PIPELINED_BOUNDED);
> {code}
> Currently, after allocate for v1,v2, we got a local partition and three
> remote partition. But actually, it should be 2 local partition and 2 remote
> partition.
> The causes of the above problems is becuase that the current allocate
> strategy is allocate the resource for execution one by one(if the execution
> can allocate from SlotGroup than get it, Otherwise ask for a new one for it).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
