[ 
http://issues.apache.org/jira/browse/HADOOP-170?page=comments#action_12376593 ] 

Yoram Arnon commented on HADOOP-170:
------------------------------------

the optimal replication factor for distributing a file in 2 hops is 
sqrt(cluster size). The worst case delivery time is 2*sqrt(n) times a single 
transfer time.

replicating to all the nodes of the dfs cluster has two disadvantages:
1. it's slower than replicating to sqrt(n), with a worst case of n.
2. if the dfs cluster is large, but multiple smaller map-reduce clusters are 
using it concurrently, there's little value in over-replication to every node 
of the cluster when it's required by a subset.

+1 on asking the job tracker for the size of the cluster and replicating based 
on that size.

> setReplication and related bug fixes
> ------------------------------------
>
>          Key: HADOOP-170
>          URL: http://issues.apache.org/jira/browse/HADOOP-170
>      Project: Hadoop
>         Type: Improvement

>   Components: fs, dfs
>     Versions: 0.1.1
>     Reporter: Konstantin Shvachko
>     Assignee: Konstantin Shvachko
>  Attachments: setReplication.patch
>
> Having variable replication (HADOOP-51) it is natural to be able to
> change replication for existing files. This patch introduces the 
> functionality.
> Here is a detailed list of issues addressed by the patch.
> 1) setReplication() and getReplication() methods are implemented.
> 2) DFSShell prints file replication for any listed file.
> 3) Bug fix. FSDirectory.delete() logs delete operation even if it is not 
> successful.
> 4) Bug fix. This is a distributed bug.
> Suppose that file replication is 3, and a client reduces it to 1.
> Two data nodes will be chosen to remove their copies, and will do that.
> After a while they will report to the name node that the copies have been 
> actually deleted.
> Until they report the name node assumes the copies still exist.
> Now the client decides to increase replication back to 3 BEFORE the data nodes
> reported the copies are deleted. Then the name node can choose one of the 
> data nodes,
> which it thinks have a block copy, to replicate the block to new data nodes.
> This setting is quite unusual but possible even without variable replications.
> 5) Logging for name and data nodes is improved in several cases.
> E.g. data nodes never logged that they deleted a block.

-- 
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators:
   http://issues.apache.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see:
   http://www.atlassian.com/software/jira

Reply via email to