[ 
https://issues.apache.org/jira/browse/DERBY-4963?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12980027#action_12980027
 ] 

Knut Anders Hatlen commented on DERBY-4963:
-------------------------------------------

I ran the single-record update performance test in derbyTesting.jar on 
Solaris/JDK6 with the write-cache disabled, and I didn't see any performance 
loss with the patch. Solaris is one of the platforms on which rws vs rwd, or 
sync(true) vs sync(false), does make a difference, so it seems you're right 
that the code path in question is not executed on recent JVMs. If there had 
been a change from one write operation to two write operations per log flush, 
the single-record update performance test should have detected it.

> Revert to FileDescriptor#sync from FileChannel#force to improve interrupt 
> resilience
> ------------------------------------------------------------------------------------
>
>                 Key: DERBY-4963
>                 URL: https://issues.apache.org/jira/browse/DERBY-4963
>             Project: Derby
>          Issue Type: Improvement
>          Components: Store
>            Reporter: Dag H. Wanvik
>            Assignee: Dag H. Wanvik
>             Fix For: 10.8.0.0
>
>         Attachments: derby-4963-1.diff, derby-4963-1.stat
>
>
> FileChannel.force is interruptable, and we really don't want to be 
> interrupted when we flush the log file.  Happily, on most platforms, we use 
> the "rws"/"rwd" file open mask which makes the writes thjemselves 
> synchronized, so no subsequent explicit file level sync is needed anyway.
> DirFile4#getRandowmAccessFile should use plain DirRandomAccessFile instead of 
> the current DirRandomAccessFile4. This will make StorageRandomAccessFile#sync 
> map to FileDescriptor#sync instead of FileChannel#force (also for NIO 
> supporting platforms). 
> Since FileDescriptor#sync does not allow synching file data only (it also 
> synchronizes metadata), those platforms which do not support write 
> synchronization will experience a performance drop, but this is the price we 
> have to pay to survive interrupts without shutting down the database on those 
> platforms. 
> Users which experience this as a problem, should update to a newer JVM which 
> does support "rws"/"rwd" in the mode argument to java.io.RandomAccessFile 
> (http://download.oracle.com/javase/6/docs/api/java/io/RandomAccessFile.html#RandomAccessFile(java.io.File,%20java.lang.String).
> Cf. also discussion on 
> https://issues.apache.org/jira/browse/DERBY-4741?focusedCommentId=12977862&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#action_12977862
>  .

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.

Reply via email to