Stephen Boesch created SPARK-2638:
-------------------------------------
Summary: Improve concurrency of fetching Map outputs
Key: SPARK-2638
URL: https://issues.apache.org/jira/browse/SPARK-2638
Project: Spark
Issue Type: Improvement
Components: Spark Core
Affects Versions: 1.0.0
Environment: All
Reporter: Stephen Boesch
Priority: Minor
Fix For: 1.1.0
This issue was noticed while perusing the MapOutputTracker source code. Notice
that the synchronization is on the containing "fetching" collection - which
makes ALL fetches wait if any fetch were occurring.
The fix is to synchronize instead on the shuffleId (interned as a string to
ensure JVM wide visibility).
def getServerStatuses(shuffleId: Int, reduceId: Int): Array[(BlockManagerId,
Long)] = {
val statuses = mapStatuses.get(shuffleId).orNull
if (statuses == null) {
logInfo("Don't have map outputs for shuffle " + shuffleId + ", fetching
them")
var fetchedStatuses: Array[MapStatus] = null
fetching.synchronized { // This is existing code
// shuffleId.toString.intern.synchronized { // New Code
if (fetching.contains(shuffleId)) {
// Someone else is fetching it; wait for them to be done
while (fetching.contains(shuffleId)) {
try {
fetching.wait()
} catch {
case e: InterruptedException =>
}
}
This is only a small code change, but the testcases to prove (a) proper
functionality and (b) proper performance improvement are not so trivial.
For (b) it is not worthwhile to add a testcase to the codebase. Instead I have
added a git project that demonstrates the concurrency/performance improvement
using the fine-grained approach . The github project is at
https://github.com/javadba/scalatesting.git . Simply run "sbt test". Note: it
is unclear how/where to include this ancillary testing/verification information
that will not be included in the git PR: i am open for any suggestions - even
as far as simply removing references to it.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
