[
https://issues.apache.org/jira/browse/FLINK-3880?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15274298#comment-15274298
]
Maximilian Michels commented on FLINK-3880:
-------------------------------------------
You're right, the synchronized map is a bottle neck. Actually, it is not even
necessary that it synchronizes. In a regular Flink job, it can only be accessed
by one task at a time. Only if the user spawned additional threads, it could be
concurrently modified. In this case the user would have to take care of the
synchronization (and if not get a ConcurrentModificationException). So we can
simply make it a normal map.
> Use ConcurrentHashMap for Accumulators
> --------------------------------------
>
> Key: FLINK-3880
> URL: https://issues.apache.org/jira/browse/FLINK-3880
> Project: Flink
> Issue Type: Improvement
> Affects Versions: 1.1.0
> Reporter: Ken Krugler
> Priority: Minor
>
> I was looking at improving DataSet performance - this is for a job created
> using the Cascading-Flink planner for Cascading 3.1.
> While doing a quick "poor man's profiler" session with one of the TaskManager
> processes, I noticed that many (most?) of the threads that were actually
> running were in this state:
> {code:java}
> "DataSource (/working1/terms) (8/20)" daemon prio=10 tid=0x00007f55673e0800
> nid=0x666a runnable [0x00007f556abcf000]
> java.lang.Thread.State: RUNNABLE
> at java.util.Collections$SynchronizedMap.get(Collections.java:2037)
> - locked <0x00000006e73fe718> (a java.util.Collections$SynchronizedMap)
> at
> org.apache.flink.api.common.functions.util.AbstractRuntimeUDFContext.getAccumulator(AbstractRuntimeUDFContext.java:162)
> at
> org.apache.flink.api.common.functions.util.AbstractRuntimeUDFContext.getLongCounter(AbstractRuntimeUDFContext.java:113)
> at
> com.dataartisans.flink.cascading.runtime.util.FlinkFlowProcess.getOrInitCounter(FlinkFlowProcess.java:245)
> at
> com.dataartisans.flink.cascading.runtime.util.FlinkFlowProcess.increment(FlinkFlowProcess.java:128)
> at
> com.dataartisans.flink.cascading.runtime.util.FlinkFlowProcess.increment(FlinkFlowProcess.java:122)
> at
> cascading.tap.hadoop.util.MeasuredRecordReader.next(MeasuredRecordReader.java:65)
> at cascading.scheme.hadoop.SequenceFile.source(SequenceFile.java:97)
> at
> cascading.tuple.TupleEntrySchemeIterator.getNext(TupleEntrySchemeIterator.java:166)
> at
> cascading.tuple.TupleEntrySchemeIterator.hasNext(TupleEntrySchemeIterator.java:139)
> at
> com.dataartisans.flink.cascading.runtime.source.TapSourceStage.readNextRecord(TapSourceStage.java:70)
> at
> com.dataartisans.flink.cascading.runtime.source.TapInputFormat.reachedEnd(TapInputFormat.java:175)
> at
> org.apache.flink.runtime.operators.DataSourceTask.invoke(DataSourceTask.java:173)
> at org.apache.flink.runtime.taskmanager.Task.run(Task.java:559)
> at java.lang.Thread.run(Thread.java:745)}}}
> {code}
> It looks like Cascading is asking Flink to increment a counter with each
> Tuple read, and that in turn is often blocked on getting access to the
> Accumulator object in a map. It looks like this is a SynchronizedMap, but
> using a ConcurrentHashMap (for example) would reduce this contention.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
