[ 
https://issues.apache.org/jira/browse/HADOOP-894?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Konstantin Shvachko updated HADOOP-894:
---------------------------------------

    Attachment: partialBlockList3.patch

In this patch:
- open takes three parameters open(src,  offset, length)
- there is an undocumented config parameter "dfs.read.prefetch.size" that 
defines the range within which 
we want all block locations to be fetch from the name-node during the open call.
- I kept 10*defaultBlockSize as the default, because 2 vs 10 does not improve 
much communication or name-node
performance, but in most cases 10 will be ALL blocks for the majority of files.
- Implemented block location caching for reads and preads.
- Included more test cases in TestPread


> dfs client protocol should allow asking for parts of the block map
> ------------------------------------------------------------------
>
>                 Key: HADOOP-894
>                 URL: https://issues.apache.org/jira/browse/HADOOP-894
>             Project: Hadoop
>          Issue Type: Improvement
>          Components: dfs
>            Reporter: Owen O'Malley
>         Assigned To: Konstantin Shvachko
>         Attachments: partialBlockList.patch, partialBlockList2.patch, 
> partialBlockList3.patch
>
>
> I think that the HDFS client protocol should change like:
> /** The meta-data about a file that was opened. */
> class OpenFileInfo {
>   /** the info for the first block */
>   public LocatedBlockInfo getBlockInfo();
>   public long getBlockSize();
>   public long getLength();
> }
> interface ClientProtocol extends VersionedProtocol {
>   public OpenFileInfo open(String name) throws IOException;
>   /** get block info for any range of blocks */
>   public LocatedBlockInfo[] getBlockInfo(String name, int blockOffset, int 
> blockLength) throws IOException;
> }
> so that the client can decide how much block info to request and when. 
> Currently, when the file is opened or an error occurs, the entire block list 
> is requested and sent.

-- 
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