[
https://issues.apache.org/jira/browse/FLINK-12852?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16877605#comment-16877605
]
zhijiang commented on FLINK-12852:
----------------------------------
The spill way could solve the deadlock issue in function, which is very similar
with previous SpillableSubpartition's behavior. If we only want to spill the
buffers between core size and max size after redistribution, it still could not
solve another existing exception of `insufficient number of network buffers'
which was always experienced in Flink for large-scale job. If we also spill
some buffers within core size to avoid that `IOException`, the performance
regression might be serious but users are not aware of it. Users might prefer
to increase buffer options to avoid performance regression if they know.
Especially for streaming job, it is better not to touch disk unless necessary.
In contrast, if we agree that the slot resource matching would be the final way
in future, then it could resolve both deadlock and `insufficient number of
network buffers' issues. And users could decide to adjust the relevant buffer
configs to make a tradeoff between performance and total resource usage. And we
might further improve the internal mechanism for decreasing the requirements of
core buffer sizes(including exclusive and core size in local buffer) to make
job still run when given limited resource.
> Deadlock occurs when requiring exclusive buffer for RemoteInputChannel
> ----------------------------------------------------------------------
>
> Key: FLINK-12852
> URL: https://issues.apache.org/jira/browse/FLINK-12852
> Project: Flink
> Issue Type: Bug
> Components: Runtime / Network
> Affects Versions: 1.7.2, 1.8.1, 1.9.0
> Reporter: Yun Gao
> Assignee: Yun Gao
> Priority: Blocker
> Labels: pull-request-available
> Fix For: 1.9.0
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> When running tests with an upstream vertex and downstream vertex, deadlock
> occurs when submitting the job:
> {code:java}
> "Sink: Unnamed (3/500)" #136 prio=5 os_prio=0 tid=0x00007f2cca81b000
> nid=0x38845 waiting on condition [0x00007f2cbe9fe000]
> java.lang.Thread.State: TIMED_WAITING (parking)
> at sun.misc.Unsafe.park(Native Method)
> - parking to wait for <0x000000073ed6b6f0> (a
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject)
> at java.util.concurrent.locks.LockSupport.parkNanos(LockSupport.java:233)
> at
> java.util.concurrent.locks.AbstractQueuedSynchronizer$ConditionObject.awaitNanos(AbstractQueuedSynchronizer.java:2078)
> at java.util.concurrent.ArrayBlockingQueue.poll(ArrayBlockingQueue.java:418)
> at
> org.apache.flink.runtime.io.network.buffer.NetworkBufferPool.requestMemorySegments(NetworkBufferPool.java:180)
> at
> org.apache.flink.runtime.io.network.buffer.NetworkBufferPool.requestMemorySegments(NetworkBufferPool.java:54)
> at
> org.apache.flink.runtime.io.network.partition.consumer.RemoteInputChannel.assignExclusiveSegments(RemoteInputChannel.java:139)
> at
> org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.assignExclusiveSegments(SingleInputGate.java:312)
> - locked <0x000000073fbc81f0> (a java.lang.Object)
> at
> org.apache.flink.runtime.io.network.partition.consumer.SingleInputGate.setup(SingleInputGate.java:220)
> at
> org.apache.flink.runtime.taskmanager.Task.setupPartionsAndGates(Task.java:836)
> at org.apache.flink.runtime.taskmanager.Task.run(Task.java:598)
> at java.lang.Thread.run(Thread.java:834)
> {code}
> This is due to the required and max of local buffer pool is not the same and
> there may be over-allocation, when assignExclusiveSegments there are no
> available memory.
>
> The detail of the scenarios is as follows: The parallelism of both upstream
> vertex and downstream vertex are 1000 and 500 respectively. There are 200 TM
> and each TM has 10696 buffers( in total and has 10 slots. For a TM that runs
> 9 upstream tasks and 1 downstream task, the 9 upstream tasks start first with
> local buffer pool \{required = 500, max = 2 * 500 + 8 = 1008}, it produces
> data quickly and each occupy about 990 buffers. Then the DownStream task
> starts and try to assigning exclusive buffers for 1500 -9 = 1491
> InputChannels. It requires 2981 buffers but only 1786 left. Since not all
> downstream tasks can start, the job will be blocked finally and no buffer can
> be released, and the deadlock finally occurred.
>
> I think although increasing the network memory solves the problem, the
> deadlock may not be acceptable. Fined grained resource management
> Flink-12761 can solve this problem, but AFAIK in 1.9 it will not include the
> network memory into the ResourceProfile.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
