flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Kruse, Sebastian" <Sebastian.Kr...@hpi.de>
Subject Re: Taskmanager memory
Date Wed, 09 Dec 2015 10:55:36 GMT
Thanks for your answers. So the problem with on-heap memory would be that the JVM would not
shrink its already allocated heap even if it is largely unused?

Pertaining to the streaming-mode: If I run Flink in that mode, can I still submit batch jobs?
Because that's what I want to do.


Thanks,

Sebastian

________________________________
From: ewenstephan@gmail.com <ewenstephan@gmail.com> on behalf of Stephan Ewen <sewen@apache.org>
Sent: Wednesday, December 9, 2015 11:15
To: user@flink.apache.org
Subject: Re: Taskmanager memory

Off heap memory is freed when the memory consuming operators release the memory.

The Java process releases that memory then on the next GC, as far as I know.

On Wed, Dec 9, 2015 at 11:01 AM, Fabian Hueske <fhueske@gmail.com<mailto:fhueske@gmail.com>>
wrote:
Streaming mode with on-heap memory won't help because the JVM allocates all memory but doesn't
convert it to managed memory internally, right?

Is offheap memory actually freed after it has been allocated as managed memory? Does this
happen after a job finishes?

2015-12-09 10:44 GMT+01:00 Stephan Ewen <sewen@apache.org<mailto:sewen@apache.org>>:
@Sebastian: Getting memory away from the JVM is tricky always, completely independent of pre-allocation
of managed memory or lazy allocation.

But here is something that may work:
  - Start Flink in streaming mode - that will make it allocate managed memory lazily
  - Set the memory to offheap memory. That way the JVM heap is small. The off-heap memory
is returned when no longer used deallocated - this releases memory much better than JVM shrinking
the heap.



On Wed, Dec 9, 2015 at 10:06 AM, Fabian Hueske <fhueske@gmail.com<mailto:fhueske@gmail.com>>
wrote:
Hi Sebastian,

There is no way to return memory from a Flink process except shutting the process down.
I think YARN could help in your setup. In a YARN setup, you can flexibly start and stop Flink
sessions with different configurations (memory, TMs, slots) or run a single job. When running
a single job, Flink will allocate resources and free them after the job is done.

Best, Fabian

2015-12-09 9:46 GMT+01:00 Kruse, Sebastian <Sebastian.Kruse@hpi.de<mailto:Sebastian.Kruse@hpi.de>>:

Hi everyone,


I am currently looking into how Flink can coexist and interoperate with other frameworks in
a cluster, such as plain single-machine processes or Spark?. ?Tachyon seems to be nice solution
to exchange data between them.


However, I think it is a problem that Flink's taskmanagers allocate their managed memory upfront
- in contrast to Spark, as far as I know. If I want ?a taskmanager to yield its main memory,
so that another process can use that memory, is there any other option besides shutting that
taskmanager down? Would it be beneficial to use YARN?

Thanks for your help!


Cheers,

Sebastian





Mime
View raw message