You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@kudu.apache.org by "Todd Lipcon (JIRA)" <ji...@apache.org> on 2017/04/24 21:44:04 UTC

[jira] [Commented] (KUDU-1726) Avoid fsync-per-block in tablet copy

    [ https://issues.apache.org/jira/browse/KUDU-1726?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15981964#comment-15981964 ] 

Todd Lipcon commented on KUDU-1726:
-----------------------------------

As a quick test, I used 'set_flag never_sync true' on a bunch of tablet servers in a cluster doing tablet copies. This increased the tablet copy throughput on the cluster from around ~400M/sec to around ~700M/sec.

> Avoid fsync-per-block in tablet copy
> ------------------------------------
>
>                 Key: KUDU-1726
>                 URL: https://issues.apache.org/jira/browse/KUDU-1726
>             Project: Kudu
>          Issue Type: Bug
>          Components: tablet
>            Reporter: Mike Percy
>
> We should be able to do a full tablet copy, keeping in mind which blocks and WAL segments have changed, and then do a bulk sync-to-disk once the full operation is complete. This would allow the kernel to schedule the IO at its leisure until durability is actually required.
> This will likely require changes from the BlockManger API and LogBlockManager implementation to support.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)