Mailing-List: contact mapreduce-user-help@hadoop.apache.org; run by ezmlm
Precedence: bulk
Reply-To: mapreduce-user@hadoop.apache.org
Received-SPF: neutral (athena.apache.org: local policy)
Message-ID: <50928B37.7030202@zfabrik.de>
Date: Thu, 01 Nov 2012 15:46:15 +0100
From: Henning Blohm <henning.blohm@zfabrik.de>
User-Agent: Mozilla/5.0 (X11; Linux x86_64;
 rv:16.0) Gecko/20121028 Thunderbird/16.0.2
MIME-Version: 1.0
To: mapreduce-user@hadoop.apache.org
Subject: Re: Virtual memory problems on Ubuntu 12.04 (a.k.a. MALLOC_ARENA_MAX
 or HADOOP-7154)
References: 
 <31615395.861.1351302327462.JavaMail.lancenorskog@Lance-Norskogs-MacBook-Pro.local>
In-Reply-To: 
 <31615395.861.1351302327462.JavaMail.lancenorskog@Lance-Norskogs-MacBook-Pro.local>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit

On 10/27/2012 03:45 AM, Lance Norskog wrote:
> 1) Java uses different variables than 'malloc'. Look up 'Java garbage collection' to find out how it all works.
What are you trying to say here? This is not a garbage collection 
problem. It's a problem with the new arena collector in glibc that has 
this impact on JVM executed multi-threaded apps (tempted to say look up 
"MALLOC_ARENA_MAX JVM virtual memory" to find out how it all works).

>
> 2) Is this a 32-bit kernel? Or Java version? Those top out at 2.1g address space. You need to run with a 64-bit kernel & Java to get real work done with Hadoop.
As said in the original post: It's all 64-bit.

-- Henning
>
> ----- Original Message -----
> | From: "Henning Blohm" <henning.blohm@zfabrik.de>
> | To: mapreduce-user@hadoop.apache.org
> | Sent: Thursday, October 25, 2012 8:52:00 AM
> | Subject: Re: Virtual memory problems on Ubuntu 12.04 (a.k.a. MALLOC_ARENA_MAX or HADOOP-7154)
> |
> | Could not get it to make sense out of MALLOC_ARENA_MAX. No .bashrc
> | etc.
> | no env script seemed to have any impact.
> |
> | Made jobs work again by setting yarn.nodemanager.vmem-pmem-ratio=10.
> | Now
> | they probably run with some obscene and unnecessary vmem allocation
> | (which I read does not come for free with the new malloc). What a
> | crappy
> | situation (and change) :-(
> |
> | Thanks,
> |    Henning
> |
> | On 10/25/2012 11:47 AM, Henning Blohm wrote:
> | > Recently I have installed data nodes on Ubuntu 12.04 and observed
> | > failing M/R jobs with errors like this:
> | >
> | > Diagnostics report from attempt_1351154628597_0002_m_000000_0:
> | > Container
> | > [pid=14529,containerID=container_1351154628597_0002_01_000002] is
> | > running beyond virtual memory limits. Current usage: 124.4mb of
> | > 1.0gb
> | > physical memory used; 2.1gb of 2.1gb virtual memory used. Killing
> | > container.
> | > Dump of the process-tree for container_1351154628597_0002_01_000002
> | > :
> | >     |- PID PPID PGRPID SESSID CMD_NAME USER_MODE_TIME(MILLIS)
> | > SYSTEM_TIME(MILLIS) VMEM_USAGE(BYTES) RSSMEM_USAGE(PAGES)
> | > FULL_CMD_LINE
> | >     |- 14529 13550 14529 14529 (java) 678 18 2265411584 31856
> | > /home/gd/gd/jdk1.6.0_35/bin/java -Djava.net.preferIPv4Stack=true
> | > -Dhadoop.metrics.log.level=WARN -Xmx1000M -XX:MaxPermSize=512M
> | > -Djava.io.tmpdir=/home/gd/gd/gi-de-nosql.cdh4-base/data/yarn/usercache/gd/appcache/application_1351154628597_0002/container_1351154628597_0002_01_000002/tmp
> | > -Dlog4j.configuration=container-log4j.properties
> | > -Dyarn.app.mapreduce.container.log.dir=/home/gd/gd/gi-de-nosql.cdh4-base/logs/application_1351154628597_0002/container_1351154628597_0002_01_000002
> | > -Dyarn.app.mapreduce.container.log.filesize=0
> | > -Dhadoop.root.logger=INFO,CLA org.apache.hadoop.mapred.YarnChild
> | > 192.168.178.25 36183 attempt_1351154628597_0002_m_000000_0 2
> | >
> | > I am using CDH4.0.1 (hadoop 2.0.0) with the Yarn M/R implementation
> | > on
> | > Ubuntu 12.04 64Bit.
> | >
> | > According to HADOOP-7154 making sure MALLOC_ARENA_MAX=1 (or 4) is
> | > exported should fix the issue.
> | >
> | > I tried the following:
> | >
> | > Exporting the environment variable MALLOC_ARENA_MAX with value 1 in
> | > all hadoop shell scrips (e.g. yarn-env.sh). Checking the
> | > launch_container.sh script that Yarn creates I can tell that it
> | > indeed
> | > contains the line
> | >
> | > export MALLOC_ARENA_MAX="1"
> | >
> | > But still I am getting the error above.
> | >
> | > In addition I tried adding
> | >
> | >     <property>
> | >        <name>mapred.child.env</name>
> | >        <value>MALLOC_ARENA_MAX=1</value>
> | >     </property>
> | >
> | > to mapred-site.xml. But that didn't seem to fix it either.
> | >
> | > Is there anything special that I need to configure on the server to
> | > make the setting effective?
> | >
> | > Any idea would be great!!
> | >
> | > Thanks,
> | >   Henning
> |
> |