Optimize locking codepath in LocalDirAllocator.getLocalPathToRead() and reduce
creating many Path objects
---------------------------------------------------------------------------------------------------------
Key: HADOOP-7016
URL: https://issues.apache.org/jira/browse/HADOOP-7016
Project: Hadoop Common
Issue Type: Bug
Components: fs
Affects Versions: 0.20.2
Reporter: Rajesh Balamohan
Fix For: 0.20.2
LocalDirAllocator.getLocalPathToRead() is called from TaskTracker to retrieve
temporary map output location. This method is synchronized and amount of time
spent on this method directly affects the performance of TaskTracker's ability
to serve faster. This JIRA is created to reduce the additional fat in
getLocalPathToRead() and reduce creating the number of Path obejcts which are
expensive.
1. getLocalPathToRead() internally calls RawLocalFileSystem.exists() in order
to check the existence of a local file. RawLocalFileSystem.exists() internally
creates a FileStatus object which needs to be populated with lastModified,
length, directory details etc. This might not be necessary for just checking
file existence.
2. Creating a Path object out of a string requires fair amount of processing.
It might be worth to store "localDirsPath" in LocalDirAllocator instead of
storing "localDirs". This would avoid the number of times Path() object is
created in LocalDirAllocator.
Any other codepath using LocalDirAllocator would also benefit from the above 2
optimizations.
The attached patch addresses the above 2 issues.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
