[
https://issues.apache.org/jira/browse/DRILL-7607?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17048164#comment-17048164
]
ASF GitHub Bot commented on DRILL-7607:
---------------------------------------
paul-rogers commented on pull request #2000: DRILL-7607: support dynamic credit
based flow control
URL: https://github.com/apache/drill/pull/2000#discussion_r385997586
##########
File path:
exec/java-exec/src/main/java/org/apache/drill/exec/work/batch/UnlimitedRawBatchBuffer.java
##########
@@ -90,14 +100,39 @@ public boolean isEmpty() {
@Override
public void add(RawFragmentBatch batch) {
+ int recordCount = batch.getHeader().getDef().getRecordCount();
+ long bathByteSize = batch.getByteCount();
+ if (recordCount != 0) {
+ //skip first header batch
+ totalBatchSize += bathByteSize;
+ sampleTimes++;
+ }
+ if (sampleTimes == maxSampleTimes) {
+ long averageBathSize = totalBatchSize / sampleTimes;
+ //make a decision
+ long limit = context.getAllocator().getLimit();
Review comment:
Unless this has been adjusted, the limit is generally meaningless; there is
no code that sets this to any reasonable value (unless something has changed
recently.) It generally defaults to 10 GB, which is far too large.
If we use the default, will each receiver try to use the 10 GB of memory?
More than 1 or 2 queries will cause OOM on many installations.
To provide a bit more background. Unless things changed recently, if the
query queue is off, there is no memory budget per query. But, if queuing is on,
there is code that allocates memory to the buffering operators (sort, join,
etc.) but not (last I looked) to exchanges. (It clearly should do such an
allocation.) So, if we use the full 10GB (or whatever value) here, the query
may far exceed the budget and the queuing mechanism will still cause the system
to exhaust memory before we hit the query limit.
In short, this memory needs to be allocated somehow as part of memory
planning. I don't see any code in this PR which will do so.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> Dynamic credit based flow control
> ---------------------------------
>
> Key: DRILL-7607
> URL: https://issues.apache.org/jira/browse/DRILL-7607
> Project: Apache Drill
> Issue Type: New Feature
> Components: Server, Execution - RPC
> Affects Versions: 1.17.0
> Reporter: Weijie Tong
> Assignee: Weijie Tong
> Priority: Major
> Fix For: 1.18.0
>
>
> Drill current has a static credit based flow control between the batch sender
> and receiver. That means ,all the sender send out their batch through the
> DataTunnel by a static 3 semaphore. To the receiver side , there's two cases,
> the UnlimitedRawBatchBuffer has a 6 * fragmentCount receiver semaphore, the
> SpoolingRawBatchBuffer acts as having unlimited receiving semaphore as it
> could flush data to disk.
> The static credit has the following weak points:
> 1. While the send batch data size is low(e.g. it has only one column bigint
> data) and the receiver has larger memory space, the sender still could not
> send out its data rapidly.
> 2. As the static credit assumption does not set the semaphore number
> according to the corresponding receiver memory space, it still have the risk
> to make the receiver OOM.
> 3. As the sender semaphore is small, it could not send its batch
> consecutively due to wait for an Ack to release one semaphore , and then ,
> the sender's corresponding execution pipeline would be halt, also the same to
> its leaf execution nodes.
> The dynamic credit based flow control could solve these problems. It starts
> from the static credit flow control. Then the receiver collects some batch
> datas to calculate the average batch size. According to the receiver side
> memory space, the receiver make a runtime sender credit and receiver side
> total credit. The receiver sends out the runtime sender credit number to the
> sender by the Ack response. The sender change to the runtime sender credit
> number when receives the Ack response with a runtime credit value.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
