[
https://issues.apache.org/jira/browse/FLINK-18991?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Roman Khachatryan updated FLINK-18991:
--------------------------------------
Description:
Curently, channel state is read not sequentially.
Inverting control would make it more efficient.
Current call chain:
{code:java}
StreamTask.readRecoveredChannelState -> strictly after reading output; when
state is *consumed*, requests partitions
SingleInputGate.readRecoveredState -> submits a task to an executor from
main task thread; loop through special channels
RecoveredInputChannel.readRecoveredState -> loop while has data;
buffer = bufferManager.requestBufferBlocking();
where bufferManager = new
BufferManager(inputGate.getMemorySegmentProvider(), this, 0);
ChannelStateReader.readInputData
{code}
Proposed call chain:
{code:java}
StreamTask.readRecoveredChannelState -> strictly after reading output; when
state is *consumed*, requests partitions
ChannelStateReader.readInputData (another thread, maybe same executor,
maybe change injection)
loop through state handles ordererd by handle, offset
request buffer from channel
pass buffer to RecoveredInputChannel.onRecoveredStateBuffer
{code}
Depends on FLINK-18989
was:
Curently, channel state is read not sequentially.
Inverting control would make it more efficient.
Current call chain:
{code:java}
StreamTask.readRecoveredChannelState -> strictly after reading output; when
state is *consumed*, requests partitions
SingleInputGate.readRecoveredState -> submits a task to an executor from
main task thread; loop through special channels
RecoveredInputChannel.readRecoveredState -> loop while has data;
buffer = bufferManager.requestBufferBlocking();
where bufferManager = new
BufferManager(inputGate.getMemorySegmentProvider(), this, 0);
ChannelStateReader.readInputData
{code}
Proposed call chain:
{code:java}
StreamTask.readRecoveredChannelState -> strictly after reading output; when
state is *consumed*, requests partitionsStreamTask.readRecoveredChannelState ->
strictly after reading output; when state is *consumed*, requests partitions
ChannelStateReader.readInputData (another thread, maybe same executor, maybe
change injection) loop through state handles ordererd by handle, offset
request buffer from channel pass buffer to
RecoveredInputChannel.onRecoveredStateBuffer
{code}
Depends on FLINK-18989
> Optimize reading InputChannel state
> -----------------------------------
>
> Key: FLINK-18991
> URL: https://issues.apache.org/jira/browse/FLINK-18991
> Project: Flink
> Issue Type: Improvement
> Components: Runtime / Task
> Reporter: Roman Khachatryan
> Assignee: Roman Khachatryan
> Priority: Major
> Fix For: 1.12.0
>
>
> Curently, channel state is read not sequentially.
> Inverting control would make it more efficient.
> Current call chain:
> {code:java}
> StreamTask.readRecoveredChannelState -> strictly after reading output; when
> state is *consumed*, requests partitions
> SingleInputGate.readRecoveredState -> submits a task to an executor from
> main task thread; loop through special channels
> RecoveredInputChannel.readRecoveredState -> loop while has data;
> buffer = bufferManager.requestBufferBlocking();
> where bufferManager = new
> BufferManager(inputGate.getMemorySegmentProvider(), this, 0);
> ChannelStateReader.readInputData
> {code}
> Proposed call chain:
> {code:java}
> StreamTask.readRecoveredChannelState -> strictly after reading output; when
> state is *consumed*, requests partitions
> ChannelStateReader.readInputData (another thread, maybe same executor,
> maybe change injection)
> loop through state handles ordererd by handle, offset
> request buffer from channel
> pass buffer to RecoveredInputChannel.onRecoveredStateBuffer
> {code}
> Depends on FLINK-18989
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
