hadoop-common-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "dhruba borthakur (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HADOOP-1003) Proposal to batch commits to edits log.
Date Fri, 09 Feb 2007 21:31:05 GMT

    [ https://issues.apache.org/jira/browse/HADOOP-1003?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12471826
] 

dhruba borthakur commented on HADOOP-1003:
------------------------------------------

We could get lot of parallelism by just moving the fsync() outside the FSNamesystem lock.
This allows all IPC threads to be writing concurrently to the edit log (kernel buffers). I
believe that an sync is pretty fast if there is no dirty data to flush. Thus, it should be
ok for all transactions to issue a sync() .

I think this might give us good parallelism without the complexity of having asynchronous
call completions on the RPC server.

> Proposal to batch commits to edits log.
> ---------------------------------------
>
>                 Key: HADOOP-1003
>                 URL: https://issues.apache.org/jira/browse/HADOOP-1003
>             Project: Hadoop
>          Issue Type: Improvement
>          Components: dfs
>            Reporter: Raghu Angadi
>         Assigned To: Sameer Paranjpye
>
> Right now most expensive namenode operations are that require commits to edits log. e.g.
creating a file, deleting, renaming etc. Most of the time is spent in fsync() of edits file
(multiple fsync() calls in the case of multiple image directories). During this time whole
namesystem is under lock and even non-mutating operations like open() are blocked.
> On a local filesystem, each fsync could take in the order of milliseconds. My understanding
is that guarantee namenode provides is that edits log is synced before replying to the client.
Without any changes to  current locking structure, I was thinking of the following for batching
multiple edits : 
>      a) a facility in RPC Server to postpone responding to a particular call (communication
with ThreadLocals may be). This is strictly not required but without it, number operations
batched would be limited to number of IPC threads.
>      b) Another Server thread that waits for pending commits to be synced and replies
back to clients. 
>      c)  fsync manager that periodically syncs the edit log and informs waiting RPCs.
The sync thread can dynamically decide to wait longer or shorter based on the load so that
we don't increase the latency when namenode is lightly loaded. Event simple policy of 'sync
if there are any mutations' will also work but that might reduce the hard disk life.
>  
> All the synchronization between these threads is a bit complicated but it can be stable.
My main concern is whether the guarantee we are providing enough for namenode operation. I
think it is enough.  
> In terms of throughput, number of creates a namenode can do should be on the same range
as number of opens it can do.
>  

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


Mime
View raw message