[
https://issues.apache.org/jira/browse/HDFS-6109?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14728538#comment-14728538
]
Young Lee commented on HDFS-6109:
---------------------------------
[~xieliang007] I was going over the impact of enabling the
'dfs.datanode.sync.behind.writes.in.background' flag and thought about
conditions under sync_file_range with SYNC_FILE_RANGE_WRITE might block. The
sync_file_range man page says exactly what you alluded to in the ticket
description:
SYNC_FILE_RANGE_WRITE
Initiate write-out of all dirty pages in the specified range
which are not presently submitted write-out. Note that even
this may block if you attempt to write more than request queue
size.
Would u please elaborate on the conditions where sync_file_range was blocking.
What were the memory, disk, number and size of the regions on the region
server, etc. I want to understand the signs to look for before enabling the
'dfs.datanode.sync.behind.writes.in.background' flag.
Thanx,
- Young
> let sync_file_range() system call run in background
> ---------------------------------------------------
>
> Key: HDFS-6109
> URL: https://issues.apache.org/jira/browse/HDFS-6109
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: datanode
> Affects Versions: 3.0.0, 2.3.0
> Reporter: Liang Xie
> Assignee: Liang Xie
> Fix For: 2.5.0
>
> Attachments: HDFS-6109-v2.txt, HDFS-6109-v3.txt, HDFS-6109-v4.txt,
> HDFS-6109.txt
>
>
> Through we passed SYNC_FILE_RANGE_WRITE to sync_file_range, to make it as
> asynchronous as possible, it still could be blocked, e.g. the os io request
> queue is full.
> Since we use sync_file_range just as a page cache advisor role:) it doesn't
> decide or guarantee the real durability, it would be nice if we could run it
> in backgroud. At least my test log showed, a few sync_file_range calls still
> cost tens of ms or more, due to the happened location is in the critical
> write path(BlockReceiver class), from a upper view, like HBase application,
> will "hung" tens of ms as well during Hlog syncing.
> Generally speaking, the patch could not improve too much, but, better than
> before, right ? :)
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
