zhuzhurk commented on a change in pull request #18376:
URL: https://github.com/apache/flink/pull/18376#discussion_r790281527
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/scheduler/SsgNetworkMemoryCalculationUtils.java
##########
@@ -148,6 +161,78 @@ private static TaskInputsOutputsDescriptor
buildTaskInputsOutputsDescriptor(
return ret;
}
+ private static Map<IntermediateDataSetID, Integer>
getMaxInputChannelNumsForDynamicGraph(
+ ExecutionJobVertex ejv) {
+
+ Map<IntermediateDataSetID, Integer> ret = new HashMap<>();
+ for (IntermediateResult consumedResult : ejv.getInputs()) {
+ ret.put(consumedResult.getId(),
getMaxInputChannelNumForResult(ejv, consumedResult));
+ }
+
+ return ret;
+ }
+
+ private static Map<IntermediateDataSetID, Integer>
getMaxSubpartitionNumsForDynamicGraph(
+ ExecutionJobVertex ejv) {
+
+ Map<IntermediateDataSetID, Integer> ret = new HashMap<>();
+
+ for (IntermediateResult intermediateResult :
ejv.getProducedDataSets()) {
+ final int maxNum =
+ Arrays.stream(intermediateResult.getPartitions())
+
.map(IntermediateResultPartition::getNumberOfSubpartitions)
+ .reduce(0, Integer::max);
+ ret.put(intermediateResult.getId(), maxNum);
+ }
+
+ return ret;
+ }
+
+ @VisibleForTesting
+ static int getMaxInputChannelNumForResult(
+ ExecutionJobVertex ejv, IntermediateResult consumedResult) {
+ DistributionPattern distributionPattern =
consumedResult.getConsumingDistributionPattern();
+
+ if (distributionPattern == DistributionPattern.ALL_TO_ALL) {
+ int numChannelsToConsumePerPartition =
+
getMaxNumOfChannelsForConsuming(consumedResult.getPartitions()[0]);
+ int numConsumedPartitions =
consumedResult.getNumberOfAssignedPartitions();
+ return numChannelsToConsumePerPartition * numConsumedPartitions;
+
+ } else if (distributionPattern == DistributionPattern.POINTWISE) {
+ int numPartitions = consumedResult.getNumberOfAssignedPartitions();
+ int numConsumers = ejv.getParallelism();
+ // when using dynamic graph, all partitions have the same number
of subpartitions
Review comment:
I took a look at IntermediateResultPartition again and yes you are
right. But I just feel that this assumption is easy to break and easy to lost
track. This is because it is more like an implicit implementation result,
rather than an explicit contract.
e.g. if later someone tried to optimized
`IntermediateResultPartition#computeNumberOfSubpartitions()` to avoid creating
extra subpartitions if the parallelism of the downstream vertex is known
already, this assumption may break. However, he/she may not be aware of the
changes here the assumption here may stay unchanged.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
