Hi guys,

we are running Hadoop 0.20.2+228, the namenode process's memory grows
overtime to occupy over 18GB.  However, if I restart the namenode, it only
occupies about 10GB after it is stable.  I am wondering if there is anyway
to figure out what is going on with namenode that causes it to grow very
rapidly and if there are any tools to make the namenode printout some
useful information on what is holding onto that memory.

Thanks,

Felix

Hi Felix,

Taking  jmap –histo:live <pid> would tell you what’s occupying the heap.

Are you using  UseConcMarkSweepGC?
If yes, and if you see bunch of
  java.net.SocksSocketImpl
  sun.nio.ch.SocketChannelImpl
from jmap histo outputs, try passing    -XX:-CMSConcurrentMTEnabled .

Background: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7113118

Koji

On 1/18/12 3:23 PM, "felix gao" <[EMAIL PROTECTED]> wrote:

Hi guys,

we are running Hadoop 0.20.2+228, the namenode process's memory grows overtime to occupy over 18GB.  However, if I restart the namenode, it only occupies about 10GB after it is stable.  I am wondering if there is anyway to figure out what is going on with namenode that causes it to grow very rapidly and if there are any tools to make the namenode printout some useful information on what is holding onto that memory.

Thanks,

Felix

Felix,

Are you noticing this from the NN Web UI or from proper JVM monitoring
tools? Just so you know, the UI had a flaw that has since been fixed:
https://issues.apache.org/jira/browse/HDFS-94

On Thu, Jan 19, 2012 at 4:53 AM, felix gao <[EMAIL PROTECTED]> wrote:
> Hi guys,
>
> we are running Hadoop 0.20.2+228, the namenode process's memory grows
> overtime to occupy over 18GB.  However, if I restart the namenode, it only
> occupies about 10GB after it is stable.  I am wondering if there is anyway
> to figure out what is going on with namenode that causes it to grow very
> rapidly and if there are any tools to make the namenode printout some useful
> information on what is holding onto that memory.
>
> Thanks,
>
> Felix

--
Harsh J
Customer Ops. Engineer, Cloudera

I noticed this from doing a top of the process and it is reporting the RES
is about 80% of the allocated.

On Wed, Jan 18, 2012 at 6:56 PM, Harsh J <[EMAIL PROTECTED]> wrote:

> Felix,
>
> Are you noticing this from the NN Web UI or from proper JVM monitoring
> tools? Just so you know, the UI had a flaw that has since been fixed:
> https://issues.apache.org/jira/browse/HDFS-94
>
> On Thu, Jan 19, 2012 at 4:53 AM, felix gao <[EMAIL PROTECTED]> wrote:
> > Hi guys,
> >
> > we are running Hadoop 0.20.2+228, the namenode process's memory grows
> > overtime to occupy over 18GB.  However, if I restart the namenode, it
> only
> > occupies about 10GB after it is stable.  I am wondering if there is
> anyway
> > to figure out what is going on with namenode that causes it to grow very
> > rapidly and if there are any tools to make the namenode printout some
> useful
> > information on what is holding onto that memory.
> >
> > Thanks,
> >
> > Felix
>
>
>
> --
> Harsh J
> Customer Ops. Engineer, Cloudera
>

Koji,

There is no Java options specified other than -Xmx24g, what are some of the
recommended options for namenode?

Thanks,

Felix

On Wed, Jan 18, 2012 at 3:39 PM, Koji Noguchi <[EMAIL PROTECTED]>wrote:

>  Hi Felix,
>
> Taking  jmap –histo:live <pid> would tell you what’s occupying the heap.
>
> Are you using  UseConcMarkSweepGC?
> If yes, and if you see bunch of
>   java.net.SocksSocketImpl
>   sun.nio.ch.SocketChannelImpl
> from jmap histo outputs, try passing    -XX:-CMSConcurrentMTEnabled .
>
> Background: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7113118
>
> Koji
>
>
>
>
> On 1/18/12 3:23 PM, "felix gao" <[EMAIL PROTECTED]> wrote:
>
> Hi guys,
>
> we are running Hadoop 0.20.2+228, the namenode process's memory grows
> overtime to occupy over 18GB.  However, if I restart the namenode, it only
> occupies about 10GB after it is stable.  I am wondering if there is anyway
> to figure out what is going on with namenode that causes it to grow very
> rapidly and if there are any tools to make the namenode printout some
> useful information on what is holding onto that memory.
>
> Thanks,
>
> Felix
>
>

Can you please check in UI, what is the heap usage. Then we can confirm whether java heap is growing or not.

top will consider native memory usage also and nio uses directByteBuffers internally.

This is good write up from Jonathan https://groups.google.com/group/asynchbase/browse_thread/thread/c45bc7ba788b2357

Regards,

Uma

________________________________
From: felix gao [[EMAIL PROTECTED]]
Sent: Friday, January 20, 2012 6:42 AM
To: [EMAIL PROTECTED]
Subject: Re: namenode grows overtime

Koji,

There is no Java options specified other than -Xmx24g, what are some of the recommended options for namenode?

Thanks,

Felix

On Wed, Jan 18, 2012 at 3:39 PM, Koji Noguchi <[EMAIL PROTECTED]<mailto:[EMAIL PROTECTED]>> wrote:
Hi Felix,

Taking  jmap –histo:live <pid> would tell you what’s occupying the heap.

Are you using  UseConcMarkSweepGC?
If yes, and if you see bunch of
  java.net.SocksSocketImpl
  sun.nio.ch.SocketChannelImpl
from jmap histo outputs, try passing    -XX:-CMSConcurrentMTEnabled .

Background: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7113118

Koji
On 1/18/12 3:23 PM, "felix gao" <[EMAIL PROTECTED]<http://[EMAIL PROTECTED]>> wrote:

Hi guys,

we are running Hadoop 0.20.2+228, the namenode process's memory grows overtime to occupy over 18GB.  However, if I restart the namenode, it only occupies about 10GB after it is stable.  I am wondering if there is anyway to figure out what is going on with namenode that causes it to grow very rapidly and if there are any tools to make the namenode printout some useful information on what is holding onto that memory.

Thanks,

Felix

Felix, if no options are specified, then I think what Harsh said makes most sense.

Koji

==============> Felix,
>
> Are you noticing this from the NN Web UI or from proper JVM monitoring
> tools? Just so you know, the UI had a flaw that has since been fixed:
> https://issues.apache.org/jira/browse/HDFS-94
> --
> Harsh J
> Customer Ops. Engineer, Cloudera

On 1/20/12 1:25 AM, "Uma Maheswara Rao G" <[EMAIL PROTECTED]> wrote:

Can you please check in UI, what is the heap usage. Then we can confirm whether java heap is growing or not.

top will consider native memory usage also and nio uses directByteBuffers internally.

This is good write up from Jonathan https://groups.google.com/group/asynchbase/browse_thread/thread/c45bc7ba788b2357

Regards,

Uma

________________________________
From: felix gao [[EMAIL PROTECTED]]
Sent: Friday, January 20, 2012 6:42 AM
To: [EMAIL PROTECTED]
Subject: Re: namenode grows overtime

Koji,

There is no Java options specified other than -Xmx24g, what are some of the recommended options for namenode?

Thanks,

Felix

On Wed, Jan 18, 2012 at 3:39 PM, Koji Noguchi <[EMAIL PROTECTED]> wrote:
Hi Felix,

Taking  jmap –histo:live <pid> would tell you what’s occupying the heap.

Are you using  UseConcMarkSweepGC?
If yes, and if you see bunch of
  java.net.SocksSocketImpl
  sun.nio.ch.SocketChannelImpl
from jmap histo outputs, try passing    -XX:-CMSConcurrentMTEnabled .

Background: http://bugs.sun.com/bugdatabase/view_bug.do?bug_id=7113118

Koji
On 1/18/12 3:23 PM, "felix gao" <[EMAIL PROTECTED] <http://[EMAIL PROTECTED]> > wrote:

Hi guys,

we are running Hadoop 0.20.2+228, the namenode process's memory grows overtime to occupy over 18GB.  However, if I restart the namenode, it only occupies about 10GB after it is stable.  I am wondering if there is anyway to figure out what is going on with namenode that causes it to grow very rapidly and if there are any tools to make the namenode printout some useful information on what is holding onto that memory.

Thanks,

Felix