Github user srowen commented on a diff in the pull request:
https://github.com/apache/spark/pull/22371#discussion_r216383808
--- Diff:
core/src/main/scala/org/apache/spark/shuffle/IndexShuffleBlockResolver.scala ---
@@ -138,13 +154,22 @@ private[spark] class IndexShuffleBlockResolver(
mapId: Int,
lengths: Array[Long],
dataTmp: File): Unit = {
+ val mapLocks = shuffleIdToLocks.get(shuffleId)
+ require(mapLocks != null, "Shuffle should be registered to
IndexShuffleBlockResolver first")
+ val lock = mapLocks.synchronized {
--- End diff --
The theory is many fewer threads would contend here because it's
per-shuffleID.
If it's an issue, then your idea of a second-level ConcurrentHashMap might
help. It's more complex than a usual Map but can allow for safe concurrent
access by a limited number of threads.
Otherwise it might be overkill as the second-level Map.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscr...@spark.apache.org
For additional commands, e-mail: reviews-h...@spark.apache.org
