Github user gczsjdy commented on a diff in the pull request:
https://github.com/apache/spark/pull/19763#discussion_r152016240
--- Diff: core/src/main/scala/org/apache/spark/MapOutputTracker.scala ---
@@ -472,16 +475,45 @@ private[spark] class MapOutputTrackerMaster(
shuffleStatuses.get(shuffleId).map(_.findMissingPartitions())
}
+ /**
+ * Try to equally divide Range(0, num) to divisor slices
+ */
+ def equallyDivide(num: Int, divisor: Int): Iterator[Seq[Int]] = {
+ assert(divisor > 0, "Divisor should be positive")
+ val (each, remain) = (num / divisor, num % divisor)
+ val (smaller, bigger) = (0 until num).splitAt((divisor-remain) * each)
+ if (each != 0) {
+ smaller.grouped(each) ++ bigger.grouped(each + 1)
+ } else {
+ bigger.grouped(each + 1)
+ }
+ }
+
/**
* Return statistics about all of the outputs for a given shuffle.
*/
def getStatistics(dep: ShuffleDependency[_, _, _]): MapOutputStatistics
= {
shuffleStatuses(dep.shuffleId).withMapStatuses { statuses =>
val totalSizes = new Array[Long](dep.partitioner.numPartitions)
- for (s <- statuses) {
- for (i <- 0 until totalSizes.length) {
- totalSizes(i) += s.getSizeForBlock(i)
+ if (statuses.length * totalSizes.length <=
+ conf.get(SHUFFLE_MAP_OUTPUT_STATISTICS_MULTITHREAD_THRESHOLD)) {
+ for (s <- statuses) {
+ for (i <- 0 until totalSizes.length) {
+ totalSizes(i) += s.getSizeForBlock(i)
+ }
+ }
+ } else {
+ val parallelism =
conf.getInt("spark.adaptive.map.statistics.cores", 8)
--- End diff --
I thought only adaptive execution code will call this. But actually it
seems after all `ShuffleMapTask`s of a stage completed this will be called,
right?
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
