osdir.com


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

回复:Memory size for network buffers


1. The max value is actually used in current calculation. If the value based on fraction * totalMem is more than max value, we take the max value as the final result.

2. I think there are many things to do behind this, such as the elastic allocation you mentioned. The network buffer is already covered as a factor in ResourceProfile which would be matched before deploying tasks. To do so, we can avoid throwing exception during task startups because the task will not be deployed into the TaskManager whose network buffer is not satisfied the task's ResourceProfile.

The current three parameters for network memory are more flexible than before that just config the extract "numberOfBuffers", so it may not be changed again if we make some improments in future .

Best,
Zhijiang
------------------------------------------------------------------
发件人:Thomas Weise <thw@xxxxxxxxxx>
发送时间:2018年11月14日(星期三) 02:37
收件人:dev <dev@xxxxxxxxxxxxxxxx>; wangzhijiang999 <wangzhijiang999@xxxxxxxxxx>
主 题:Re: Memory size for network buffers

Hi Zhijiang,

Thanks for taking a look. Please note that the configuration isn't real, it
was just for black box testing of these configuration parameters. Since the
max parameter is never used, so why does it even exist? Currently memory
size is always fixed, either based on min or fraction.

Curious if there was a larger plan behind this? I don't see why automatic
and elastic allocation cannot be provided. Required network buffers could
be allocated as needed, instead of throwing an exception.

Thanks!


On Mon, Nov 12, 2018 at 11:39 PM zhijiang
<wangzhijiang999@xxxxxxxxxx.invalid> wrote:

> Hi Thomas,
>
> The current calculation for network buffer size is
> "Math.min(taskmanager.network.memory.max,
> Math.max(taskmanager.network.memory.min, fraction * totalMem))".
> Based on your below configuration, the result is just 32768 bytes (8
> buffers) from taskmanager.network.memory.min.
> If you want to config the fixed network buffers, you can set the same
> values for min and max parameters, ignore the fraction value in the
> configuration.
> BTW, you can set the unit for these paramenters, such as min : 32kb.
>
> As for the thoughts of min and max setting,  it is difficult to exactly
> know how many network buffers are needed in TaskManager startup and which
> kinds of tasks would be deployed to run in this TaskManager later.  For
> example, the batch jobs can make use of as many network buffers as system
> can spare. But for stream jobs, the spare buffers can be used for other
> places for possible improvements. In order to keep the possibility of
> future improments and not change the configuration setting, we retain these
> current parameters.
>
> For you job I think you should increase the min value for more network
> buffers, the current 4 buffers are indeed not enough for common jobs.
>
> Best,
> Zhijiang
> ------------------------------------------------------------------
> 发件人:Thomas Weise <thw@xxxxxxxxxx>
> 发送时间:2018年11月13日(星期二) 13:11
> 收件人:dev <dev@xxxxxxxxxxxxxxxx>
> 主 题:Memory size for network buffers
>
> Hi,
>
> I'm trying to understand the intention behind the size parameters for
> network buffers, specifically max, min and fraction. The reason we are
> looking at it is an attempt to make the memory allocation elastic, so that
> memory is allocated according to the actual number of buffers required
> (within a range), without the need to tune this for every deployment.
>
> As of Flink 1.5, there are 3 parameters, but they all result in a fixed
> allocation, which is not what we were looking for.
>
> Here is an example just to illustrate it:
>
> taskmanager.network.memory.fraction: 0.000001
> taskmanager.network.memory.min: 32768
> taskmanager.network.memory.max: 1073741824
> taskmanager.memory.segment-size: 8192
>
> I wanted fraction to be out of the picture (but 0 isn't an acceptable
> value).
>
> Then set min to something tiny that my job will exceed and max to something
> too large to reach. Unfortunately, that fails:
>
> java.io.IOException: Insufficient number of network buffers: required 8,
> but only 0 available. The total number of network buffers is currently set
> to 4 of 8192 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'.
>
> So the question then is, why have min and max? Or is the intention to have
> a different implementation in the future?
>
> Thanks,
> Thomas
>
>