[
https://issues.apache.org/jira/browse/HDFS-2421?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jing Zhao updated HDFS-2421:
----------------------------
Attachment: performance_SerialNumberManager.png
TestSerialNumberManager.java
HDFS-2421.patch
Write a simple test program to compare the two programs: ConcurrentHashMap vs.
HashMap + synchronized. The TestSerialNumberManager.java is the test code. The
ConcurrentHashMap version changes the original synchronization semantic, but
seems fine if every thread is generating serial number for different names.
A simple performance measurement is included (results are in milliseconds). For
both read and write, every thread is doing
10000 sets of r/w operations, and I tested scenarios with 10, 20, 50, and 100
threads. The test is on a MAC with 8G memory and a 2.2 GHz i7 processor.
> Improve the concurrency of SerialNumberMap in NameNode
> -------------------------------------------------------
>
> Key: HDFS-2421
> URL: https://issues.apache.org/jira/browse/HDFS-2421
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: name-node
> Reporter: Hairong Kuang
> Assignee: Weiyan Wang
> Fix For: 0.24.0
>
> Attachments: HDFS-2421.patch, performance_SerialNumberManager.png,
> TestSerialNumberManager.java
>
>
> After enabled permission checking in our HDFS test cluster, our benchmark
> observed a significant reduced concurrency in NameNode. Investigation showed
> that most threads were blocked at acquiring the lock of
> org.apache.hadoop.hdfs.server.namenode.SerialNumberManager$SerialNumberMap.
> We used concurrentHashMap to replace Hashmap + synchronized methods, which
> greatly improved the situation.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
