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 -
>    0.48) = 53 GB
>    - On Yarn: jvmHeap = (total - Max(cutoff-min, total * cutoff-ratio)) *
>    (1 - networkFraction) = (102GB - Max(600MB, 102GB * 0.25)) * (1 - 0.48) =
>    40.6GB
>
>
Are you running Flink on Mesos? I think Flink has not automatically set
-Xmx on Mesos.


BTW, from your screenshot the physical memory is 123GB, so 1/4 of that is
much closer to 29GB if we consider there are some rounding errors and
accuracy loss.


Thank you~

Xintong Song



On Fri, Jun 12, 2020 at 4:33 PM Vijay Balakrishnan <bvija...@gmail.com>
wrote:

> 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 about the 29GB
> figure.
>
> On Thu, Jun 11, 2020 at 9:14 PM Xintong Song <tonysong...@gmail.com>
> wrote:
>
>> 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 month.
>>
>> Regarding your questions,
>>
>>    - "Physical Memory" displayed on the web ui stands for the total
>>    memory on your machine. This information is retrieved from your OS. It is
>>    not related to the network memory calculation. It is displayed mainly for
>>    historical reasons.
>>    - The error message means that you have about 26.8 GB network memory
>>    (877118 * 32768 bytes), and your job is trying to use more.
>>    - The "total memory" referred in network memory calculation is:
>>       - jvm-heap + network, if managed memory is configured on-heap
>>       (default)
>>          - According to your screenshot, the managed memory
>>          on-heap/off-heap configuration is not touched, so this should be 
>> your case.
>>       - jvm-heap + managed + network, if managed memory is configured
>>       off-heap
>>    - The network memory size is actually derived reversely. Flink reads
>>    the max heap size from JVM (and the managed memory size from configuration
>>    if it is configured off-heap), and derives the network memory size with 
>> the
>>    following equation.
>>       - networkMem = Min(networkMax, Max(networkMin, jvmMaxHeap /
>>       (1-networkFraction) * networkFraction))
>>       - In your case, networkMem = Min(50GB, Max(500MB, 29GB / (1-0.48)
>>       * 0.48)) = 26.8GB
>>
>> One thing I don't understand is, why do you only have 29GB heap size when
>> "taskmanager.heap.size" is configured to be "1044221m" (about 102 GB). The
>> JVM heap size ("-Xmx" & "-Xms") is calculated as follows. I'll use "total"
>> to represent "taskmanager.heap.size" for short. Also omitted the
>> calculations when managed memory is configured off-heap.
>>
>>    - Standalone: jvmHeap = total * (1 - networkFraction) = 102 GB * (1 -
>>    0.48) = 53 GB
>>    - On Yarn: jvmHeap = (total - Max(cutoff-min, total * cutoff-ratio))
>>    * (1 - networkFraction) = (102GB - Max(600MB, 102GB * 0.25)) * (1 - 0.48) 
>> =
>>    40.6GB
>>
>> Have you specified a custom "-Xmx" parameter?
>>
>> Thank you~
>>
>> Xintong Song
>>
>>
>>
>> On Fri, Jun 12, 2020 at 7:50 AM Vijay Balakrishnan <bvija...@gmail.com>
>> wrote:
>>
>>> Hi,
>>> Get this error:
>>> java.io.IOException: Insufficient number of network buffers: required 2,
>>> but only 0 available. The total number of network buffers is currently set
>>> to 877118 of 32768 bytes each. You can increase this number by setting the
>>> configuration keys 'taskmanager.network.memory.fraction',
>>> 'taskmanager.network.memory.min', and 'taskmanager.network.memory.max'.
>>> akka.pattern.AskTimeoutException: Ask timed out on
>>> [Actor[akka://flink/user/dispatcher#-1420732632]] after [10000 ms]. Message
>>> of type [org.apache.flink.runtime.rpc.messages.LocalFencedMessage]. A
>>> typical reason for `AskTimeoutException` is that the recipient actor didn't
>>> send a reply.
>>>
>>>
>>> Followed docs here:
>>>
>>> https://ci.apache.org/projects/flink/flink-docs-release-1.9/ops/mem_setup.html
>>>
>>> network = Min(max, Max(min, fraction x total)  //what does Total mean -
>>> The max JVM heap is used to derive the total memory for the calculation of
>>> network buffers. - can I see it in the Flink Dashboard ??? 117GB here ?
>>> = Min(50G, Max(500mb, Max(0.48 * 117G))  ) = MIn(50G, 56.16G)= 50G
>>> 877118 of 32768 bytes each comes to 28.75GB. So, why is it failing ?
>>> Used this in flink-conf.yaml:
>>>     taskmanager.numberOfTaskSlots: 10
>>>     rest.server.max-content-length: 314572800
>>>     taskmanager.network.memory.fraction: 0.45
>>>     taskmanager.network.memory.max: 50gb
>>>     taskmanager.network.memory.min: 500mb
>>>     akka.ask.timeout: 240s
>>>     cluster.evenly-spread-out-slots: true
>>>     akka.tcp.timeout: 240s
>>> taskmanager.network.request-backoff.initial: 5000
>>> taskmanager.network.request-backoff.max: 30000
>>> web.timeout:1000000
>>> web.refresh-interval:6000
>>>
>>> Saw some old calc about buffers
>>> (slots/Tm * slots/TM) * #TMs * 4
>>> =10 * 10 * 47 * 4 = 18,800 buffers.
>>>
>>> What am I missing in the network buffer calc ??
>>>
>>> TIA,
>>>
>>>
>>>

Reply via email to