[jira] Commented: (HADOOP-1565) DFSScalability: reduce memory usage of namenode

["Konstantin Shvachko (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-1565?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12516838 ] 

Konstantin Shvachko commented on HADOOP-1565:
---------------------------------------------

- I agree ArrayList should better serve the purpose than the TreeMap.
It saves us about 50 bytes per directory entry according to my calculations.
- hashcode. I don't think waisting 4 bytes per INode plus the complexity of supporting the hash code oriented
 ordering worth the performance gain we get from that. I would compare names as they are same as we did before.
 We  are talking about 10-20 entries per directory and file names of length 10 on average.
- is there a reason for reimplementing binary search rather than using Arrays.binarySearch()?
- children = new ArrayList<INode>(5);
   5 should be a constant
- System.out.println() should be removed.


> DFSScalability: reduce memory usage of namenode
> -----------------------------------------------
>
>                 Key: HADOOP-1565
>                 URL: https://issues.apache.org/jira/browse/HADOOP-1565
>             Project: Hadoop
>          Issue Type: Bug
>            Reporter: dhruba borthakur
>            Assignee: dhruba borthakur
>         Attachments: memoryReduction2.patch
>
>
> Experiments have demonstrated that a single file/block needs about 300 to 500 bytes of main memory on a 64-bit Namenode. This puts some limitations on the size of the file system that a single namenode can support. Most of this overhead occurs because a block and/or filename is inserted into multiple TreeMaps and/or HashSets.
> Here are a few ideas that can be measured to see if an appreciable reduction of memory usage occurs:
> 1. Change FSDirectory.children from a TreeMap to an array. Do binary search in this array while looking up children. This saves a TreeMap object for every intermediate node in the directory tree.
> 2. Change INode from an inner class. This saves on one "parent object" reference for each INODE instance. 4 bytes per inode.
> 3. Keep all DatanodeDescriptors in an array. BlocksMap.nodes[] is currently a 64-bit reference to the DatanodeDescriptor object. Instead, it can be a 'short'. This will probably save about 16 bytes per block.
> 4. Change DatanodeDescriptor.blocks from a SortedTreeMap to a HashMap? Block report processing CPU cost can increase.
> For the records: TreeMap has the following fields:
> 	Object key;
> 	Object value;
> 	Entry left = null;
> 	Entry right = null;
> 	Entry parent;
> 	boolean color = BLACK;
> and HashMap object:
> 	final Object key;
> 	Object value;
> 	final int hash;
> 	Entry next;

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