[
https://issues.apache.org/jira/browse/FLINK-7456?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16266984#comment-16266984
]
ASF GitHub Bot commented on FLINK-7456:
---------------------------------------
Github user NicoK commented on a diff in the pull request:
https://github.com/apache/flink/pull/4552#discussion_r152985049
--- Diff:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/NetworkEnvironment.java
---
@@ -224,17 +224,9 @@ public void registerTask(Task task) throws IOException
{
BufferPool bufferPool = null;
try {
- if
(gate.getConsumedPartitionType().isCreditBased()) {
- // Create a fixed-size buffer
pool for floating buffers and assign exclusive buffers to input channels
directly
- bufferPool =
networkBufferPool.createBufferPool(extraNetworkBuffersPerGate,
extraNetworkBuffersPerGate);
-
gate.assignExclusiveSegments(networkBufferPool, networkBuffersPerChannel);
- } else {
- int maxNumberOfMemorySegments =
gate.getConsumedPartitionType().isBounded() ?
-
gate.getNumberOfInputChannels() * networkBuffersPerChannel +
-
extraNetworkBuffersPerGate : Integer.MAX_VALUE;
- bufferPool =
networkBufferPool.createBufferPool(gate.getNumberOfInputChannels(),
-
maxNumberOfMemorySegments);
- }
+ // Create a fixed-size buffer pool for
floating buffers and assign exclusive buffers to input channels directly
+ bufferPool =
networkBufferPool.createBufferPool(extraNetworkBuffersPerGate,
extraNetworkBuffersPerGate);
+
gate.assignExclusiveSegments(networkBufferPool, networkBuffersPerChannel);
--- End diff --
What about the non-bounded partition type that we use for batch processing?
Shouldn't we use an unbounded number of floating buffers there, as previously?
> Implement Netty sender incoming pipeline for credit-based
> ---------------------------------------------------------
>
> Key: FLINK-7456
> URL: https://issues.apache.org/jira/browse/FLINK-7456
> Project: Flink
> Issue Type: Sub-task
> Components: Network
> Reporter: zhijiang
> Assignee: zhijiang
> Fix For: 1.5.0
>
>
> This is a part of work for credit-based network flow control.
> On sender side, each subpartition view maintains an atomic integer
> {{currentCredit}} from receiver. Once receiving the messages of
> {{PartitionRequest}} and {{AddCredit}}, the {{currentCredit}} is added by
> deltas.
> Each view also maintains an atomic boolean field to mark it as registered
> available for transfer to make sure it is enqueued in handler only once. If
> the {{currentCredit}} increases from zero and there are available buffers in
> the subpartition, the corresponding view will be enqueued for transferring
> data.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
