[
https://issues.apache.org/jira/browse/FLINK-1435?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14287340#comment-14287340
]
Ufuk Celebi commented on FLINK-1435:
------------------------------------
Thanks for reporting the issue.
I agree that there should be an informative error message. Currently, the
memory manager is initialized from the config value and tries to allocate all
memory as byte buffers. At some point it just fails, when not enough memory is
available. In the GlobalBufferPool, we actually catch and rethrow the
OutOfMemoryException and log the missing amount of memory. We should add
something similar as well or put the memory manager instantiation into a try
catch block.
Would you like to provide a patch for this?
> TaskManager does not log missing memory error on start up
> ---------------------------------------------------------
>
> Key: FLINK-1435
> URL: https://issues.apache.org/jira/browse/FLINK-1435
> Project: Flink
> Issue Type: Bug
> Components: TaskManager
> Affects Versions: 0.7.0-incubating
> Reporter: Malte Schwarzer
> Priority: Minor
> Labels: memorymanager
>
> When using bin/start-cluster.sh to start TaskManagers and a worker node is
> failing to start because of missing memory, you do not receive any error
> messages in log files.
> Worker node has only 15000M memory available, but it is configured with
> Maximum heap size: 40000 MiBytes. Task manager does not join the cluster.
> Process hangs.
> Last lines of log looks like this:
> ...
> ... - - Starting with 12 incoming and 12 outgoing connection threads.
> ... - Setting low water mark to 16384 and high water mark to 32768 bytes.
> ... - Instantiated PooledByteBufAllocator with direct arenas: 24, heap
> arenas: 0, page size (bytes): 65536, chunk size (bytes): 16777216.
> ... - Using 0.7 of the free heap space for managed memory.
> ... - Initializing memory manager with 24447 megabytes of memory. Page size
> is 32768 bytes.
> (END)
> Error message about not enough memory is missing.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
