hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Arpit Agarwal (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-8722) Optimize datanode writes for small writes and flushes
Date Fri, 10 Jul 2015 19:56:06 GMT

    [ https://issues.apache.org/jira/browse/HDFS-8722?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14622825#comment-14622825

Arpit Agarwal commented on HDFS-8722:

Nice catch. A quick test shows ~15% throughput regression since 2.6.0 for \[small writes+hflush\].
||Release||IO Size||Throughput (KBps)||

Chunk aligned writes are at par with 2.6.0.
||Release||IO Size||Throughput (KBps)||

1GB file with 128 MB blocks.

> Optimize datanode writes for small writes and flushes
> -----------------------------------------------------
>                 Key: HDFS-8722
>                 URL: https://issues.apache.org/jira/browse/HDFS-8722
>             Project: Hadoop HDFS
>          Issue Type: Improvement
>    Affects Versions: 2.7.1
>            Reporter: Kihwal Lee
>            Assignee: Kihwal Lee
>            Priority: Critical
>         Attachments: HDFS-8722.patch, HDFS-8722.v1.patch
> After the data corruption fix by HDFS-4660, the CRC recalculation for partial chunk is
executed more frequently, if the client repeats writing few bytes and calling hflush/hsync.
 This is because the generic logic forces CRC recalculation if on-disk data is not CRC chunk
aligned. Prior to HDFS-4660, datanode blindly accepted whatever CRC client provided, if the
incoming data is chunk-aligned. This was the source of the corruption.
> We can still optimize for the most common case where a client is repeatedly writing small
number of bytes followed by hflush/hsync with no pipeline recovery or append, by allowing
the previous behavior for this specific case.  If the incoming data has a duplicate portion
and that is at the last chunk-boundary before the partial chunk on disk, datanode can use
the checksum supplied by the client without redoing the checksum on its own.  This reduces
disk reads as well as CPU load for the checksum calculation.
> If the incoming packet data goes back further than the last on-disk chunk boundary, datanode
will still do a recalculation, but this occurs rarely during pipeline recoveries. Thus the
optimization for this specific case should be sufficient to speed up the vast majority of

This message was sent by Atlassian JIRA

View raw message