hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Chris Trezzo (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HDFS-8791) block ID-based DN storage layout can be very slow for datanode on ext4
Date Thu, 26 Nov 2015 01:50:11 GMT

    [ https://issues.apache.org/jira/browse/HDFS-8791?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15027993#comment-15027993

Chris Trezzo commented on HDFS-8791:

Thanks all for the comments.

bq. how long did these upgrades take in the 0.5 million, 1.2 million, and 2.7 million block
I have attached a [new version|https://issues.apache.org/jira/secure/attachment/12774454/32x32DatanodeLayoutTesting-v2.pdf]
of the testing document with more details around the upgrade testing. The upgrade for the
above case was a setup with very low block density and the data node upgraded with a startup
time of 1 minute. I did do an upgrade test with a data node that had around 2 million blocks
in total. In that case the hard linking alone took around 9 minutes.

[~andrew.wang] Agreed. I think it would be awesome if we could get this into branch-2 and
I am currently in the process of adding unit tests.

[~kihwal] I took a look at a node during upgrade, and it seemed like the {{previous.tmp}}
directory did indeed have the old layout like it should. Maybe I am misunderstanding which
directory you are looking at, so I will continue to investigate.

As a side note: I am still early in my search, but I can't seem to find where in a unit test
we actually verify that the {{finalized}} directory does indeed have the correct layout after
an upgrade. The same goes for if the {{previous.tmp}} directory actually has the old format
during an upgrade that isn't finalized yet. I see {{TestDatanodeLayoutUpgrade#testUpgradeToIdBasedLayout}},
but a {{null}} verifier is passed in to the {{upgradeAndVerify}} method. Additionally, all
of the other tests (i.e. TestDFSFinalize, TestRollingUpgradeRollback, TestRollingUpgrade)
seem to either be layout agnostic or simply check the {{VERSION}} file. I will continue to

> block ID-based DN storage layout can be very slow for datanode on ext4
> ----------------------------------------------------------------------
>                 Key: HDFS-8791
>                 URL: https://issues.apache.org/jira/browse/HDFS-8791
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: datanode
>    Affects Versions: 2.6.0, 2.8.0, 2.7.1
>            Reporter: Nathan Roberts
>            Assignee: Chris Trezzo
>            Priority: Critical
>         Attachments: 32x32DatanodeLayoutTesting-v1.pdf, 32x32DatanodeLayoutTesting-v2.pdf,
> We are seeing cases where the new directory layout causes the datanode to basically cause
the disks to seek for 10s of minutes. This can be when the datanode is running du, and it
can also be when it is performing a checkDirs(). Both of these operations currently scan all
directories in the block pool and that's very expensive in the new layout.
> The new layout creates 256 subdirs, each with 256 subdirs. Essentially 64K leaf directories
where block files are placed.
> So, what we have on disk is:
> - 256 inodes for the first level directories
> - 256 directory blocks for the first level directories
> - 256*256 inodes for the second level directories
> - 256*256 directory blocks for the second level directories
> - Then the inodes and blocks to store the the HDFS blocks themselves.
> The main problem is the 256*256 directory blocks. 
> inodes and dentries will be cached by linux and one can configure how likely the system
is to prune those entries (vfs_cache_pressure). However, ext4 relies on the buffer cache to
cache the directory blocks and I'm not aware of any way to tell linux to favor buffer cache
pages (even if it did I'm not sure I would want it to in general).
> Also, ext4 tries hard to spread directories evenly across the entire volume, this basically
means the 64K directory blocks are probably randomly spread across the entire disk. A du type
scan will look at directories one at a time, so the ioscheduler can't optimize the corresponding
seeks, meaning the seeks will be random and far. 
> In a system I was using to diagnose this, I had 60K blocks. A DU when things are hot
is less than 1 second. When things are cold, about 20 minutes.
> How do things get cold?
> - A large set of tasks run on the node. This pushes almost all of the buffer cache out,
causing the next DU to hit this situation. We are seeing cases where a large job can cause
a seek storm across the entire cluster.
> Why didn't the previous layout see this?
> - It might have but it wasn't nearly as pronounced. The previous layout would be a few
hundred directory blocks. Even when completely cold, these would only take a few a hundred
seeks which would mean single digit seconds.  
> - With only a few hundred directories, the odds of the directory blocks getting modified
is quite high, this keeps those blocks hot and much less likely to be evicted.

This message was sent by Atlassian JIRA

View raw message