Github user mengxr commented on a diff in the pull request:
https://github.com/apache/spark/pull/1124#discussion_r13951639
--- Diff: core/src/main/scala/org/apache/spark/executor/Executor.scala ---
@@ -212,7 +208,14 @@ private[spark] class Executor(
val serializedDirectResult = ser.serialize(directResult)
logInfo("Serialized size of result for " + taskId + " is " +
serializedDirectResult.limit)
val serializedResult = {
- if (serializedDirectResult.limit >= akkaFrameSize - 1024) {
+ // Akka's message frame size. If task result is bigger than
this, we use the block manager
+ // to send the result back.
+ // TODO: [SPARK-1112] We use the min frame size to determine
whether to use Akka to send
+ // TODO: the task result or block manager. Since this is via the
backend, whose actor
+ // TODO: system is initialized before receiving the Spark conf,
and hence it does not know
+ // TODO: `spark.akka.frameSize`. A temporary solution is using
the min frame size.
+ // [SPARK-2156] We subtract 200K to leave some space for other
data in the Akka message.
+ if (serializedDirectResult.limit >= AkkaUtils.minFrameSizeBytes
- 200 * 1024) {
--- End diff --
I tested with a double array of size very close to 10 * 1024 * 1024. The
akka message overhead is about 30-60K. This PR doesn't fix the issues with
receiving new tasks from the driver that are bigger than 10MB. @pwendell is
working on it.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
