hadoop-hdfs-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Matthew Byng-Maddick (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HDFS-1379) Multihoming brokenness in HDFS
Date Tue, 07 Sep 2010 17:11:33 GMT

    [ https://issues.apache.org/jira/browse/HDFS-1379?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12906880#action_12906880

Matthew Byng-Maddick commented on HDFS-1379:

Hi Allen,

Thanks for your comments - I looked for other things first, and I didn't notice HADOOP-6364
- thanks for linking it!

I'm aware that multi-homing is known not to work, but it seems to me that being able to define
the bind address and then having the namenode give out something that it has constructed using
incomplete information to hdfs clients is more than just a simple "multi-homing is broken"

In response to the points:
a) I'm well aware of that - segregating the traffic with our other network is important as
a lot of people have access to that network, but keeping a cluster network which is separate
is sensible - it means it can be provisioned differently with potentially different QoS at
switches etc. - We're not just doing IPC/RPC on that - but the entire datanode<->datanode
and client<->datanode comms too - as our clients run within that network - ideally I'd
have the jobtracker bind to INADDR_ANY (IN6ADDR_ANY) so you could submit a job from anywhere,
but all the datanode comms work - but that isn't possible because of the dual use of mapred.job.tracker.
b) correct, though what's also exposed on the (more) public side is bits of JMX, rather than
just trivial availability monitoring, and that does help to point out what hadoop is seeing.
At some stage I'd like to have availability monitoring on both sides, but that's not possible
right now.

> Multihoming brokenness in HDFS
> ------------------------------
>                 Key: HDFS-1379
>                 URL: https://issues.apache.org/jira/browse/HDFS-1379
>             Project: Hadoop HDFS
>          Issue Type: Bug
>          Components: hdfs client, name-node
>    Affects Versions: 0.20.1
>         Environment: Multi-homed namenode and datanodes. hadoop-0.20.1 (cloudera distribution
on linux)
>            Reporter: Matthew Byng-Maddick
> We have a setup where - because we only have a very few machines (4 x 16 core) we're
looking at co-locating namenodes and datanodes. We also have front-end and back-end networks.
Set-up is something like:
> * machine1
> ** front-end
> ** back-end
> * machine2
> ** front-end
> ** back-end
> * machine3
> ** front-end
> ** back-end
> * machine4
> ** front-end
> ** back-end
> On each, the property *slave.host.name* is configured with the 192. address, (the *.dns.interface
settings don't actually seem to help, but that's a separate problem), and the *dfs.datanode.address*
is bound to the 192.168.24.x address on :50010, similarly the *dfs.datanode.ipc.address* is
bound there.
> In order to get efficient use of our machines, we bring up a namenode on one of them
(this then rsyncs the latest namenode fsimage etc) by bringing up a VIP on each side (we use
the 10.18.80.x side for monitoring, rather than actual hadoop comms), and binding the namenode
to that - on the inside this is
> The namenode now knows about 4 datanodes - These datanodes know
how they're bound. However, when the datanode is telling an external hdfs client where to
store the blocks, it gives out as one of the addresses (despite the datanode
not being bound there) - because that's where the datanode->namenode RPC comes from.
> This is wrong because if you've bound the datanode explicitly (using *dfs.datanode.address*)
then that's should be the only address the namenode can give out (it's reasonable, given your
comms model not to support NAT between clients and data slaves). If you bind it to * then
your normal rules for slave.host.name, dfs.datanode.dns.interface etc should take precedence.
> This may already be fixed in later releases than 0.20.1 - but if it isn't it should probably
be - you explicitly allow binding of the datanode addresses - it's unreasonable to expect
that comms to the datanode will always come from those addresses - especially in multi-homed
environments (and separating traffic out by network - especially when dealing with large volumes
of data) is useful.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message