Github user ksakellis commented on a diff in the pull request:
https://github.com/apache/spark/pull/4067#discussion_r23950220
--- Diff: core/src/main/scala/org/apache/spark/CacheManager.scala ---
@@ -47,9 +49,13 @@ private[spark] class CacheManager(blockManager:
BlockManager) extends Logging {
val inputMetrics = blockResult.inputMetrics
val existingMetrics = context.taskMetrics
.getInputMetricsForReadMethod(inputMetrics.readMethod)
- existingMetrics.addBytesRead(inputMetrics.bytesRead)
+ existingMetrics.incBytesRead(inputMetrics.bytesRead)
- new InterruptibleIterator(context,
blockResult.data.asInstanceOf[Iterator[T]])
+ val iter = blockResult.data.asInstanceOf[Iterator[T]]
+ new InterruptibleIterator(context,
AfterNextInterceptingIterator(iter, (next: T) => {
+ existingMetrics.incRecordsRead(1)
--- End diff --
@sryza right. So how do you propose we increment the bytes and records read
in a threadsafe way? If we use a @volatile Long we can't safely do an increment
unless we guarantee that only one thread is accessing the InputMetrics at any
one time. I guess this is an okay assumption now but doesn't that open
ourselves up to race conditions down the line when we add more multithreading?
Looking at:
http://stackoverflow.com/questions/2538070/atomic-operation-cost it doesn't
seem like the cost of CAS is that high, there is at most 2 cacheline misses for
this integer and only 1 if other CPUs are not reading and writing from it. Am
i misinterpreting this?
---
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.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
