[
https://issues.apache.org/jira/browse/FLINK-15963?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
zhijiang updated FLINK-15963:
-----------------------------
Description:
On sender side while writing the BufferResponse message, it always allocates
the new direct ByteBuf from netty allocator to write header part for every
message.
Maybe we can maintain a fixed ByteBuf per channel and make use of it to write
header part for all the BufferResponse messages, then verify how it effects the
performance/memory overhead in e2e/benchmark.
This improvement is also related to another argue that whether it is feasible
to suspend writing when the socket is not writable at the moment. If the
socket is not writable temporarily, the server handler would continue writing &
flushing BufferResponse messages until reaching the maximum watermark in netty
stack. And only the header ByteBuf would occupy the netty watermark at the
moment.
was:
On sender side while writing the BufferResponse message, it always allocates
the new direct ByteBuf from netty allocator to write header part for every
message.
Maybe we can make use of a fixed ByteBuf to write header part for all the
BufferResponse messages, then verify how it effects the performance/memory
overhead in e2e/benchmark.
> Reuse the same ByteBuf while writing the BufferResponse header
> --------------------------------------------------------------
>
> Key: FLINK-15963
> URL: https://issues.apache.org/jira/browse/FLINK-15963
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Network
> Reporter: zhijiang
> Priority: Major
>
> On sender side while writing the BufferResponse message, it always allocates
> the new direct ByteBuf from netty allocator to write header part for every
> message.
> Maybe we can maintain a fixed ByteBuf per channel and make use of it to write
> header part for all the BufferResponse messages, then verify how it effects
> the performance/memory overhead in e2e/benchmark.
> This improvement is also related to another argue that whether it is feasible
> to suspend writing when the socket is not writable at the moment. If the
> socket is not writable temporarily, the server handler would continue writing
> & flushing BufferResponse messages until reaching the maximum watermark in
> netty stack. And only the header ByteBuf would occupy the netty watermark at
> the moment.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
