[
https://issues.apache.org/jira/browse/FLINK-9913?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16595110#comment-16595110
]
ASF GitHub Bot commented on FLINK-9913:
---------------------------------------
zhijiangW commented on a change in pull request #6417: [FLINK-9913][runtime]
Improve output serialization only once in RecordWriter
URL: https://github.com/apache/flink/pull/6417#discussion_r213343581
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/api/writer/RecordWriter.java
##########
@@ -87,62 +86,71 @@ public RecordWriter(ResultPartitionWriter writer,
ChannelSelector<T> channelSele
this.numChannels = writer.getNumberOfSubpartitions();
- /*
- * The runtime exposes a channel abstraction for the produced
results
- * (see {@link ChannelSelector}). Every channel has an
independent
- * serializer.
- */
- this.serializers = new SpanningRecordSerializer[numChannels];
+ this.serializer = new SpanningRecordSerializer<T>();
this.bufferBuilders = new Optional[numChannels];
for (int i = 0; i < numChannels; i++) {
- serializers[i] = new SpanningRecordSerializer<T>();
bufferBuilders[i] = Optional.empty();
}
}
public void emit(T record) throws IOException, InterruptedException {
+ serializer.serializeRecord(record);
+
for (int targetChannel : channelSelector.selectChannels(record,
numChannels)) {
- sendToTarget(record, targetChannel);
+ copyToTarget(targetChannel);
}
+
+ // Make sure we don't hold onto the large intermediate
serialization buffer for too long
+ serializer.prune();
}
/**
* This is used to broadcast Streaming Watermarks in-band with records.
This ignores
* the {@link ChannelSelector}.
*/
public void broadcastEmit(T record) throws IOException,
InterruptedException {
+ serializer.serializeRecord(record);
+
for (int targetChannel = 0; targetChannel < numChannels;
targetChannel++) {
- sendToTarget(record, targetChannel);
+ copyToTarget(targetChannel);
}
+
+ serializer.prune();
}
/**
* This is used to send LatencyMarks to a random target channel.
*/
public void randomEmit(T record) throws IOException,
InterruptedException {
- sendToTarget(record, rng.nextInt(numChannels));
- }
+ serializer.serializeRecord(record);
- private void sendToTarget(T record, int targetChannel) throws
IOException, InterruptedException {
- RecordSerializer<T> serializer = serializers[targetChannel];
+ copyToTarget(rng.nextInt(numChannels));
- SerializationResult result = serializer.addRecord(record);
+ serializer.prune();
+ }
+ private void copyToTarget(int targetChannel) throws IOException,
InterruptedException {
+ // We should reset the initial position of the intermediate
serialization buffer before
+ // copying, so the serialization results can be copied to
multiple target buffers.
+ serializer.reset();
+
+ BufferBuilder bufferBuilder = getBufferBuilder(targetChannel);
+ SerializationResult result =
serializer.copyToBufferBuilder(bufferBuilder);
while (result.isFullBuffer()) {
- if (tryFinishCurrentBufferBuilder(targetChannel,
serializer)) {
- // If this was a full record, we are done. Not
breaking
- // out of the loop at this point will lead to
another
- // buffer request before breaking out (that
would not be
- // a problem per se, but it can lead to stalls
in the
- // pipeline).
- if (result.isFullRecord()) {
- break;
- }
+ tryFinishCurrentBufferBuilder(targetChannel);
Review comment:
You pointed out a good question!
1. Considering `tryFinishCurrentBufferBuilder()`, the logic is somewhat
different from before. In the past, the buffer builder may be empty when
calling `tryFinishCurrentBufferBuilder()`, then it returns a boolean value to
indicate the result. But now, we know the buffer builder is always present when
calling `tryFinishCurrentBufferBuilder`, so we may change it to
`finishCurrentBufferBuilder()` seems more appropriate. And adds the check code
instead as following:
```
private void finishCurrentBufferBuilder(int targetChannel) {
checkState(bufferBuilders[targetChannel].isPresent());
BufferBuilder bufferBuilder =
bufferBuilders[targetChannel].get();
bufferBuilders[targetChannel] = Optional.empty();
numBytesOut.inc(bufferBuilder.finish());
}
```
2. But I think we still need call `finishCurrentBufferBuilder()` after
checking every full buffer. Because in the while loop the serialized record may
be copied into several different buffer builders, and each buffer build
referenced by the `bufferBuilders ` needs to be finished when full.
3. It actually seems better than before, as the loop exit not rely on the
return value of `tryFinishCurrentBufferBuilder()`.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> Improve output serialization only once in RecordWriter
> ------------------------------------------------------
>
> Key: FLINK-9913
> URL: https://issues.apache.org/jira/browse/FLINK-9913
> Project: Flink
> Issue Type: Improvement
> Components: Network
> Affects Versions: 1.6.0
> Reporter: zhijiang
> Assignee: zhijiang
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.7.0
>
>
> Currently the {{RecordWriter}} emits output into multi channels via
> {{ChannelSelector}} or broadcasts output to all channels directly. Each
> channel has a separate {{RecordSerializer}} for serializing outputs, that
> means the output will be serialized as many times as the number of selected
> channels.
> As we know, data serialization is a high cost operation, so we can get good
> benefits by improving the serialization only once.
> I would suggest the following changes for realizing it.
> # Only one {{RecordSerializer}} is created in {{RecordWriter}} for all the
> channels.
> # The output is serialized into the intermediate data buffer only once for
> different channels.
> # The intermediate serialization results are copied into different
> {{BufferBuilder}}s for different channels.
> An additional benefit by using a single serializer for all channels is that
> we get a potentially significant reduction on heap space overhead from fewer
> intermediate serialization buffers (only once we got over 5MiB, these buffers
> were pruned back to 128B!).
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
