Hello Arvid Heise,
Thanks for replying! Based on your suggestion, I put together the following
snippet for the config:
val config = new Configuration()
private val newMemorySize =
config.get(TaskManagerOptions.MEMORY_SEGMENT_SIZE)
.multiply(4)
.multiply(config.get(TaskManagerOptions.
Hi Abelm,
you can calculate the required number of buffers as follows:
- You need to know the number of network connections which depend on the
network shuffles (ignoring rescaling and forward channels here, but they
require less buffers).
- If you have s slots per taskmanager and parallelism p, y
Hi Thilo,
the number of required network buffers depends on your data exchanges and
parallelism.
For each shuffling data exchange (what you need for group by), you ideally
have #slots-per-TM^2 * #TMs * 4 buffers.
So I'm assuming you have 11 machines and 8 slots per machine. Then for best
performa
Hi Schneider,
The error message suggests that your task managers are not configured with
enough network memory. You would need to increase the network memory
configuration. See this doc [1] for more details.
Thank you~
Xintong Song
[1]
https://ci.apache.org/projects/flink/flink-docs-release-1.
>
> Assuming no -Xmx is set, the doc above says 1/4 of physical memory i.e
> 29GB will be used.
>
This is true.
So, if I can set env.java.opts: "-Xmx102g" in flink-conf.yaml, I am
> assuming the heap max of 102Gb will be used in the N/w mem calculation.
> Is that the right way to set env.java.opt
Hi Xintong,
Just to be clear. I haven't set any -Xmx -i will check our scripts again.
Assuming no -Xmx is set, the doc above says 1/4 of physical memory i.e 29GB
will be used.
So, if I can set env.java.opts: "-Xmx102g" in flink-conf.yaml, I am
assuming the heap max of 102Gb will be used in the N/w
Flink should have calculated the heap size and set the -Xms, according to
the equations I mentioned. So if you haven't set an customized -Xmx that
overwrites this, it should not use the default 1.4 of physical memory.
>
>
>- Standalone: jvmHeap = total * (1 - networkFraction) = 102 GB * (1 -
>
Thx, Xintong for a great answer. Much appreciated.
https://ci.apache.org/projects/flink/flink-docs-release-1.9/ops/mem_setup.html#jvm-heap
Max heap: if -Xmx is set then it is its value else ΒΌ of physical machine
memory estimated by the JVM
No -Xmx is set.So, 1/4 of 102GB = 25.5GB but not sure a
Hi Vijay,
The memory configurations in Flink 1.9 and previous versions are indeed
complicated and confusing. That is why we made significant changes to it in
Flink 1.10. If possible, I would suggest upgrading to Flink 1.10, or the
upcoming Flink 1.11 which is very likely to be released in this mon
Yes, you are right, the exception was caused as task managers were heavily
loaded. I checked ganglia metrics and CPU usage was very high. I reduced
parallelism and ran with 5000 buffers and didn't get any exception.
Thanks,
Tarandeep
On Tue, May 3, 2016 at 2:19 AM, Ufuk Celebi wrote:
> Hey Tara
Hey Tarandeep,
I think the failures are unrelated. Regarding the number of network
buffers:
https://ci.apache.org/projects/flink/flink-docs-release-1.0/setup/config.html#configuring-the-network-buffers
The timeouts might occur, because the task managers are pretty loaded.
I would suggest to incr
Hi Guido,
please check this section of the configuration documentation:
https://ci.apache.org/projects/flink/flink-docs-release-0.10/setup/config.html#configuring-the-network-buffers
It should answer your questions. Please let us know, if not.
Cheers, Fabian
2015-11-27 16:41 GMT+01:00 Guido :
12 matches
Mail list logo
