[
https://issues.apache.org/jira/browse/HDFS-3228?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13962029#comment-13962029
]
Colin Patrick McCabe commented on HDFS-3228:
--------------------------------------------
It's worth noting that BlockReaderLocal already implements a form of readahead,
in that it reads and checksums {{dfs.client.read.shortcircuit.buffer.size}}
bytes at a time. Typically this is 1 MB. (Since HDFS-5634, we will also read
fewer bytes than {{dfs.client.read.shortcircuit.buffer.size}} if readahead is
turned down or turned off.)
Of course, this 1 MB read is done in the foreground (you don't get back any
bytes until it finishes). An fadvise-based system would operate a little bit
differently.
> Use fadvise in local read path
> ------------------------------
>
> Key: HDFS-3228
> URL: https://issues.apache.org/jira/browse/HDFS-3228
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: hdfs-client, performance
> Reporter: Henry Robinson
> Assignee: Henry Robinson
>
> The read path through BlockReaderLocal does not take advantage of readahead
> or drop-behind in the way that BlockSender does. We could arguably stand to
> gain even more from hinting about read patterns to the kernel here, so we
> should add the same mechanisms to BlockReaderLocal.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
