MapReduce, mail # user - Re: Insight on why distcp becomes slower when adding nodemanager

Insight on why distcp becomes slower when adding nodemanager

Alexandre Fouche 2012-10-29, 15:12

Hi,

Can someone give some insight on why a "distcp" of 600 files of a few hundred bytes from s3n:// to local hdfs is taking 46s when using a yarn-nodemanager EC2 instance with 16GB memory (which by the way i think is jokingly long), and taking 3mn30s when adding a second yarn-nodemanager (a small instance with 1.7GB memory) ?
I would have expected it to be a bit faster, not 5xlonger !

I have the same issue when i stop the small instance nodemanager and restart it to join the processing after the big nodemanager instance was already submitted the job.

I am using Cloudera latest Yarn+HDFS on Amazon (rebranded Centos 6)

   #Staging 14:58:04 root@datanode2:hadoop-yarn: rpm -qa |grep hadoop
   hadoop-hdfs-datanode-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   hadoop-mapreduce-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   hadoop-0.20-mapreduce-0.20.2+1261-1.cdh4.1.1.p0.4.el6.x86_64
   hadoop-yarn-nodemanager-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   hadoop-mapreduce-historyserver-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   hadoop-hdfs-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   hadoop-client-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   hadoop-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   hadoop-yarn-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
   #Staging 14:39:51 root@resourcemanager:hadoop-yarn: HADOOP_MAPRED_HOME=/usr/lib/hadoop-mapreduce time hadoop distcp -overwrite s3n://xxx:[EMAIL PROTECTED]ev/* hdfs:///tmp/something/a

   12/10/29 14:40:12 INFO tools.DistCp: Input Options: DistCpOptions{atomicCommit=false, syncFolder=false, deleteMissing=false, ignoreFailures=false, maxMaps=20, sslConfigurationFile='null', copyStrategy='uniformsize', sourceFileListing=null, sourcePaths=[s3n://xxx:[EMAIL PROTECTED]ev/*], targetPath=hdfs:/tmp/something/a}
   12/10/29 14:40:18 WARN conf.Configuration: io.sort.mb is deprecated. Instead, use mapreduce.task.io.sort.mb
   12/10/29 14:40:18 WARN conf.Configuration: io.sort.factor is deprecated. Instead, use mapreduce.task.io.sort.factor
   12/10/29 14:40:19 INFO mapreduce.JobSubmitter: number of splits:15
   12/10/29 14:40:19 WARN conf.Configuration: mapred.jar is deprecated. Instead, use mapreduce.job.jar
   12/10/29 14:40:19 WARN conf.Configuration: mapred.map.tasks.speculative.execution is deprecated. Instead, use mapreduce.map.speculative
   12/10/29 14:40:19 WARN conf.Configuration: mapred.reduce.tasks is deprecated. Instead, use mapreduce.job.reduces
   12/10/29 14:40:19 WARN conf.Configuration: mapred.mapoutput.value.class is deprecated. Instead, use mapreduce.map.output.value.class
   12/10/29 14:40:19 WARN conf.Configuration: mapreduce.map.class is deprecated. Instead, use mapreduce.job.map.class
   12/10/29 14:40:19 WARN conf.Configuration: mapred.job.name is deprecated. Instead, use mapreduce.job.name
   12/10/29 14:40:19 WARN conf.Configuration: mapreduce.inputformat.class is deprecated. Instead, use mapreduce.job.inputformat.class
   12/10/29 14:40:19 WARN conf.Configuration: mapred.output.dir is deprecated. Instead, use mapreduce.output.fileoutputformat.outputdir
   12/10/29 14:40:19 WARN conf.Configuration: mapreduce.outputformat.class is deprecated. Instead, use mapreduce.job.outputformat.class
   12/10/29 14:40:19 WARN conf.Configuration: mapred.map.tasks is deprecated. Instead, use mapreduce.job.maps
   12/10/29 14:40:19 WARN conf.Configuration: mapred.mapoutput.key.class is deprecated. Instead, use mapreduce.map.output.key.class
   12/10/29 14:40:19 WARN conf.Configuration: mapred.working.dir is deprecated. Instead, use mapreduce.job.working.dir
   12/10/29 14:40:20 INFO mapred.ResourceMgrDelegate: Submitted application application_1351504801306_0015 to ResourceManager at resourcemanager.cwsdev.cleverscale.com/10.60.106.130:8032
   12/10/29 14:40:20 INFO mapreduce.Job: The url to track the job: http://ip-10-60-106-130.ec2.internal:8088/proxy/application_1351504801306_0015/
   12/10/29 14:40:20 INFO tools.DistCp: DistCp job-id: job_1351504801306_0015
   12/10/29 14:40:20 INFO mapreduce.Job: Running job: job_1351504801306_0015
   12/10/29 14:40:27 INFO mapreduce.Job: Job job_1351504801306_0015 running in uber mode : false
   12/10/29 14:40:27 INFO mapreduce.Job: map 0% reduce 0%
   12/10/29 14:40:42 INFO mapreduce.Job: map 6% reduce 0%
   12/10/29 14:40:43 INFO mapreduce.Job: map 33% reduce 0%
   12/10/29 14:40:44 INFO mapreduce.Job: map 40% reduce 0%
   12/10/29 14:40:48 INFO mapreduce.Job: map 46% reduce 0%
   12/10/29 14:43:04 INFO mapreduce.Job: map 56% reduce 0%
   12/10/29 14:43:05 INFO mapreduce.Job: map 58% reduce 0%
   12/10/29 14:43:08 INFO mapreduce.Job: map 62% reduce 0%
   12/10/29 14:43:09 INFO mapreduce.Job: map 68% reduce 0%
   12/10/29 14:43:15 INFO mapreduce.Job: map 75% reduce 0%
   12/10/29 14:43:16 INFO mapreduce.Job: map 82% reduce 0%
   12/10/29 14:43:25 INFO mapreduce.Job: map 85% reduce 0%
   12/10/29 14:43:26 INFO mapreduce.Job: map 87% reduce 0%
   12/10/29 14:43:29 INFO mapreduce.Job: map 90% reduce 0%
   12/10/29 14:43:35 INFO mapreduce.Job: map 93% reduce 0%
   12/10/29 14:43:37 INFO mapreduce.Job: map 96% reduce 0%
   12/10/29 14:43:40 INFO mapreduce.Job: map 100% reduce 0%
   12/10/29 14:43:40 INFO mapreduce.Job: Job job_1351504801306_0015 completed successfully
   12/10/29 14:43:40 INFO mapreduce.Job: Counters: 35
   File System Counters
   FILE: Number of bytes read=1800
   FILE: Number of bytes written=1050895
   FILE: Number of read operations=0
   FILE: Number of large read operations=0
   FILE: Number of write operations=0
   HDFS: Number of bytes read=22157
   HDFS: Number of bytes written=101379
   HDFS: Number of read operations=519
   HDFS: Number of large read operations=0
   HDFS: Number of write operations=201
   S3N: Number of bytes read=101379
   S3N: Number of bytes written=0
   S3N: Number of read operations=0
   S3N: Number of l

Re: Insight on why distcp becomes slower when adding nodemanager

Harsh J 2012-10-29, 16:30

On your second low-memory NM instance, did you ensure to lower the
yarn.nodemanager.resource.memory-mb property specifically to avoid
swapping due to excessive resource grants? The default offered is 8 GB
(>> 1.7 GB you have).

On Mon, Oct 29, 2012 at 8:42 PM, Alexandre Fouche
<[EMAIL PROTECTED]> wrote:
> Hi,
>
> Can someone give some insight on why a "distcp" of 600 files of a few
> hundred bytes from s3n:// to local hdfs is taking 46s when using a
> yarn-nodemanager EC2 instance with 16GB memory (which by the way i think is
> jokingly long), and taking 3mn30s when adding a second yarn-nodemanager (a
> small instance with 1.7GB memory) ?
> I would have expected it to be a bit faster, not 5xlonger !
>
> I have the same issue when i stop the small instance nodemanager and restart
> it to join the processing after the big nodemanager instance was already
> submitted the job.
>
> I am using Cloudera latest Yarn+HDFS on Amazon (rebranded Centos 6)
>
> #Staging 14:58:04 root@datanode2:hadoop-yarn: rpm -qa |grep hadoop
> hadoop-hdfs-datanode-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> hadoop-mapreduce-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> hadoop-0.20-mapreduce-0.20.2+1261-1.cdh4.1.1.p0.4.el6.x86_64
> hadoop-yarn-nodemanager-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> hadoop-mapreduce-historyserver-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> hadoop-hdfs-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> hadoop-client-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> hadoop-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> hadoop-yarn-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>
>
> #Staging 14:39:51 root@resourcemanager:hadoop-yarn:
> HADOOP_MAPRED_HOME=/usr/lib/hadoop-mapreduce time hadoop distcp -overwrite
> s3n://xxx:[EMAIL PROTECTED]ev/* hdfs:///tmp/something/a
>
> 12/10/29 14:40:12 INFO tools.DistCp: Input Options:
> DistCpOptions{atomicCommit=false, syncFolder=false, deleteMissing=false,
> ignoreFailures=false, maxMaps=20, sslConfigurationFile='null',
> copyStrategy='uniformsize', sourceFileListing=null,
> sourcePaths=[s3n://xxx:[EMAIL PROTECTED]ev/*],
> targetPath=hdfs:/tmp/something/a}
> 12/10/29 14:40:18 WARN conf.Configuration: io.sort.mb is deprecated.
> Instead, use mapreduce.task.io.sort.mb
> 12/10/29 14:40:18 WARN conf.Configuration: io.sort.factor is deprecated.
> Instead, use mapreduce.task.io.sort.factor
> 12/10/29 14:40:19 INFO mapreduce.JobSubmitter: number of splits:15
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.jar is deprecated.
> Instead, use mapreduce.job.jar
> 12/10/29 14:40:19 WARN conf.Configuration:
> mapred.map.tasks.speculative.execution is deprecated. Instead, use
> mapreduce.map.speculative
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.reduce.tasks is
> deprecated. Instead, use mapreduce.job.reduces
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.mapoutput.value.class
> is deprecated. Instead, use mapreduce.map.output.value.class
> 12/10/29 14:40:19 WARN conf.Configuration: mapreduce.map.class is
> deprecated. Instead, use mapreduce.job.map.class
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.job.name is
> deprecated. Instead, use mapreduce.job.name
> 12/10/29 14:40:19 WARN conf.Configuration: mapreduce.inputformat.class
> is deprecated. Instead, use mapreduce.job.inputformat.class
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.output.dir is
> deprecated. Instead, use mapreduce.output.fileoutputformat.outputdir
> 12/10/29 14:40:19 WARN conf.Configuration: mapreduce.outputformat.class
> is deprecated. Instead, use mapreduce.job.outputformat.class
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.map.tasks is
> deprecated. Instead, use mapreduce.job.maps
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.mapoutput.key.class is
> deprecated. Instead, use mapreduce.map.output.key.class
> 12/10/29 14:40:19 WARN conf.Configuration: mapred.working.dir is
> deprecated. Instead, use mapreduce.job.working.dir

Harsh J

Re: Insight on why distcp becomes slower when adding nodemanager

Michael Segel 2012-10-29, 19:04

how many times did you test it?

need to rule out aberrations.

On Oct 29, 2012, at 11:30 AM, Harsh J <[EMAIL PROTECTED]> wrote:

> On your second low-memory NM instance, did you ensure to lower the
> yarn.nodemanager.resource.memory-mb property specifically to avoid
> swapping due to excessive resource grants? The default offered is 8 GB
> (>> 1.7 GB you have).
>
> On Mon, Oct 29, 2012 at 8:42 PM, Alexandre Fouche
> <[EMAIL PROTECTED]> wrote:
>> Hi,
>>
>> Can someone give some insight on why a "distcp" of 600 files of a few
>> hundred bytes from s3n:// to local hdfs is taking 46s when using a
>> yarn-nodemanager EC2 instance with 16GB memory (which by the way i think is
>> jokingly long), and taking 3mn30s when adding a second yarn-nodemanager (a
>> small instance with 1.7GB memory) ?
>> I would have expected it to be a bit faster, not 5xlonger !
>>
>> I have the same issue when i stop the small instance nodemanager and restart
>> it to join the processing after the big nodemanager instance was already
>> submitted the job.
>>
>> I am using Cloudera latest Yarn+HDFS on Amazon (rebranded Centos 6)
>>
>> #Staging 14:58:04 root@datanode2:hadoop-yarn: rpm -qa |grep hadoop
>> hadoop-hdfs-datanode-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>> hadoop-mapreduce-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>> hadoop-0.20-mapreduce-0.20.2+1261-1.cdh4.1.1.p0.4.el6.x86_64
>> hadoop-yarn-nodemanager-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>> hadoop-mapreduce-historyserver-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>> hadoop-hdfs-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>> hadoop-client-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>> hadoop-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>> hadoop-yarn-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>
>>
>> #Staging 14:39:51 root@resourcemanager:hadoop-yarn:
>> HADOOP_MAPRED_HOME=/usr/lib/hadoop-mapreduce time hadoop distcp -overwrite
>> s3n://xxx:[EMAIL PROTECTED]ev/* hdfs:///tmp/something/a
>>
>> 12/10/29 14:40:12 INFO tools.DistCp: Input Options:
>> DistCpOptions{atomicCommit=false, syncFolder=false, deleteMissing=false,
>> ignoreFailures=false, maxMaps=20, sslConfigurationFile='null',
>> copyStrategy='uniformsize', sourceFileListing=null,
>> sourcePaths=[s3n://xxx:[EMAIL PROTECTED]ev/*],
>> targetPath=hdfs:/tmp/something/a}
>> 12/10/29 14:40:18 WARN conf.Configuration: io.sort.mb is deprecated.
>> Instead, use mapreduce.task.io.sort.mb
>> 12/10/29 14:40:18 WARN conf.Configuration: io.sort.factor is deprecated.
>> Instead, use mapreduce.task.io.sort.factor
>> 12/10/29 14:40:19 INFO mapreduce.JobSubmitter: number of splits:15
>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.jar is deprecated.
>> Instead, use mapreduce.job.jar
>> 12/10/29 14:40:19 WARN conf.Configuration:
>> mapred.map.tasks.speculative.execution is deprecated. Instead, use
>> mapreduce.map.speculative
>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.reduce.tasks is
>> deprecated. Instead, use mapreduce.job.reduces
>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.mapoutput.value.class
>> is deprecated. Instead, use mapreduce.map.output.value.class
>> 12/10/29 14:40:19 WARN conf.Configuration: mapreduce.map.class is
>> deprecated. Instead, use mapreduce.job.map.class
>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.job.name is
>> deprecated. Instead, use mapreduce.job.name
>> 12/10/29 14:40:19 WARN conf.Configuration: mapreduce.inputformat.class
>> is deprecated. Instead, use mapreduce.job.inputformat.class
>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.output.dir is
>> deprecated. Instead, use mapreduce.output.fileoutputformat.outputdir
>> 12/10/29 14:40:19 WARN conf.Configuration: mapreduce.outputformat.class
>> is deprecated. Instead, use mapreduce.job.outputformat.class
>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.map.tasks is
>> deprecated. Instead, use mapreduce.job.maps
>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.mapoutput.key.class is

Re: Insight on why distcp becomes slower when adding nodemanager

Alexandre Fouche 2012-10-31, 12:37

These instances have no swap. I tried 5 or 6 times in a row, and modified the yarn.nodemanager.resource.memory-mb but it did not help. Later on, i'll replace the openjdk with the Oracle java SE 1.6.31 to see if it improves overall performance.
Now i am running everything on medium instances for prototyping, and while this is better, i still find it abusively slow. Maybe bad hadoop performance on less than xlarge memory instances is to be expected on EC2 ?

--
Alexandre Fouche
Lead operations engineer, cloud architect
http://www.cleverscale.com | @cleverscale
Sent with Sparrow (http://www.sparrowmailapp.com/?sig)
On Monday 29 October 2012 at 20:04, Michael Segel wrote:

> how many times did you test it?
>
> need to rule out aberrations.
>
> On Oct 29, 2012, at 11:30 AM, Harsh J <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
>
> > On your second low-memory NM instance, did you ensure to lower the
> > yarn.nodemanager.resource.memory-mb property specifically to avoid
> > swapping due to excessive resource grants? The default offered is 8 GB
> > (>> 1.7 GB you have).
> >
> > On Mon, Oct 29, 2012 at 8:42 PM, Alexandre Fouche
> > <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
> > > Hi,
> > >
> > > Can someone give some insight on why a "distcp" of 600 files of a few
> > > hundred bytes from s3n:// to local hdfs is taking 46s when using a
> > > yarn-nodemanager EC2 instance with 16GB memory (which by the way i think is
> > > jokingly long), and taking 3mn30s when adding a second yarn-nodemanager (a
> > > small instance with 1.7GB memory) ?
> > > I would have expected it to be a bit faster, not 5xlonger !
> > >
> > > I have the same issue when i stop the small instance nodemanager and restart
> > > it to join the processing after the big nodemanager instance was already
> > > submitted the job.
> > >
> > > I am using Cloudera latest Yarn+HDFS on Amazon (rebranded Centos 6)
> > >
> > > #Staging 14:58:04 root@datanode2:hadoop-yarn: rpm -qa |grep hadoop
> > > hadoop-hdfs-datanode-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > hadoop-mapreduce-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > hadoop-0.20-mapreduce-0.20.2+1261-1.cdh4.1.1.p0.4.el6.x86_64
> > > hadoop-yarn-nodemanager-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > hadoop-mapreduce-historyserver-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > hadoop-hdfs-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > hadoop-client-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > hadoop-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > hadoop-yarn-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > >
> > >
> > > #Staging 14:39:51 root@resourcemanager:hadoop-yarn:
> > > HADOOP_MAPRED_HOME=/usr/lib/hadoop-mapreduce time hadoop distcp -overwrite
> > > s3n://xxx:[EMAIL PROTECTED]ev (mailto:[EMAIL PROTECTED]ev)/* hdfs:///tmp/something/a
> > >
> > > 12/10/29 14:40:12 INFO tools.DistCp: Input Options:
> > > DistCpOptions{atomicCommit=false, syncFolder=false, deleteMissing=false,
> > > ignoreFailures=false, maxMaps=20, sslConfigurationFile='null',
> > > copyStrategy='uniformsize', sourceFileListing=null,
> > > sourcePaths=[s3n://xxx:[EMAIL PROTECTED]ev (mailto:[EMAIL PROTECTED]ev)/*],
> > > targetPath=hdfs:/tmp/something/a}
> > > 12/10/29 14:40:18 WARN conf.Configuration: io.sort.mb is deprecated.
> > > Instead, use mapreduce.task.io.sort.mb
> > > 12/10/29 14:40:18 WARN conf.Configuration: io.sort.factor is deprecated.
> > > Instead, use mapreduce.task.io.sort.factor
> > > 12/10/29 14:40:19 INFO mapreduce.JobSubmitter: number of splits:15
> > > 12/10/29 14:40:19 WARN conf.Configuration: mapred.jar is deprecated.
> > > Instead, use mapreduce.job.jar
> > > 12/10/29 14:40:19 WARN conf.Configuration:
> > > mapred.map.tasks.speculative.execution is deprecated. Instead, use
> > > mapreduce.map.speculative
> > > 12/10/29 14:40:19 WARN conf.Configuration: mapred.reduce.tasks is
> > > deprecated. Instead, use mapreduce.job.reduces
> > > 12/10/29 14:40:19 WARN conf.Configuration: mapred.mapoutput.value.class

Re: Insight on why distcp becomes slower when adding nodemanager

Michael Segel 2012-10-31, 19:23

Not sure.

Lots of things can effect your throughput.
Networking is my first guess. Which is why I asked about the number of times you've run the same test to see if there is a wide variation in timings.

On Oct 31, 2012, at 7:37 AM, Alexandre Fouche <[EMAIL PROTECTED]> wrote:

> These instances have no swap. I tried 5 or 6 times in a row, and modified the yarn.nodemanager.resource.memory-mb but it did not help. Later on, i'll replace the openjdk with the Oracle java SE 1.6.31 to see if it improves overall performance.
> Now i am running everything on medium instances for prototyping, and while this is better, i still find it abusively slow. Maybe bad hadoop performance on less than xlarge memory instances is to be expected on EC2 ?
>
>
> --
> Alexandre Fouche
> Lead operations engineer, cloud architect
> http://www.cleverscale.com | @cleverscale
> Sent with Sparrow
>
> On Monday 29 October 2012 at 20:04, Michael Segel wrote:
>
>> how many times did you test it?
>>
>> need to rule out aberrations.
>>
>> On Oct 29, 2012, at 11:30 AM, Harsh J <[EMAIL PROTECTED]> wrote:
>>
>>> On your second low-memory NM instance, did you ensure to lower the
>>> yarn.nodemanager.resource.memory-mb property specifically to avoid
>>> swapping due to excessive resource grants? The default offered is 8 GB
>>> (>> 1.7 GB you have).
>>>
>>> On Mon, Oct 29, 2012 at 8:42 PM, Alexandre Fouche
>>> <[EMAIL PROTECTED]> wrote:
>>>> Hi,
>>>>
>>>> Can someone give some insight on why a "distcp" of 600 files of a few
>>>> hundred bytes from s3n:// to local hdfs is taking 46s when using a
>>>> yarn-nodemanager EC2 instance with 16GB memory (which by the way i think is
>>>> jokingly long), and taking 3mn30s when adding a second yarn-nodemanager (a
>>>> small instance with 1.7GB memory) ?
>>>> I would have expected it to be a bit faster, not 5xlonger !
>>>>
>>>> I have the same issue when i stop the small instance nodemanager and restart
>>>> it to join the processing after the big nodemanager instance was already
>>>> submitted the job.
>>>>
>>>> I am using Cloudera latest Yarn+HDFS on Amazon (rebranded Centos 6)
>>>>
>>>> #Staging 14:58:04 root@datanode2:hadoop-yarn: rpm -qa |grep hadoop
>>>> hadoop-hdfs-datanode-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>> hadoop-mapreduce-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>> hadoop-0.20-mapreduce-0.20.2+1261-1.cdh4.1.1.p0.4.el6.x86_64
>>>> hadoop-yarn-nodemanager-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>> hadoop-mapreduce-historyserver-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>> hadoop-hdfs-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>> hadoop-client-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>> hadoop-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>> hadoop-yarn-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>
>>>>
>>>> #Staging 14:39:51 root@resourcemanager:hadoop-yarn:
>>>> HADOOP_MAPRED_HOME=/usr/lib/hadoop-mapreduce time hadoop distcp -overwrite
>>>> s3n://xxx:[EMAIL PROTECTED]ev/* hdfs:///tmp/something/a
>>>>
>>>> 12/10/29 14:40:12 INFO tools.DistCp: Input Options:
>>>> DistCpOptions{atomicCommit=false, syncFolder=false, deleteMissing=false,
>>>> ignoreFailures=false, maxMaps=20, sslConfigurationFile='null',
>>>> copyStrategy='uniformsize', sourceFileListing=null,
>>>> sourcePaths=[s3n://xxx:[EMAIL PROTECTED]ev/*],
>>>> targetPath=hdfs:/tmp/something/a}
>>>> 12/10/29 14:40:18 WARN conf.Configuration: io.sort.mb is deprecated.
>>>> Instead, use mapreduce.task.io.sort.mb
>>>> 12/10/29 14:40:18 WARN conf.Configuration: io.sort.factor is deprecated.
>>>> Instead, use mapreduce.task.io.sort.factor
>>>> 12/10/29 14:40:19 INFO mapreduce.JobSubmitter: number of splits:15
>>>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.jar is deprecated.
>>>> Instead, use mapreduce.job.jar
>>>> 12/10/29 14:40:19 WARN conf.Configuration:
>>>> mapred.map.tasks.speculative.execution is deprecated. Instead, use
>>>> mapreduce.map.speculative
>>>> 12/10/29 14:40:19 WARN conf.Configuration: mapred.reduce.tasks is
>>>> deprecated. Instead, use mapreduce.job.reduces

Re: Insight on why distcp becomes slower when adding nodemanager

Marcos Ortiz 2012-10-31, 20:27

On 10/31/2012 02:23 PM, Michael Segel wrote:
> Not sure.
>
> Lots of things can effect your throughput.
> Networking is my first guess. Which is why I asked about the number of
> times you've run the same test to see if there is a wide variation in
> timings.
>
> On Oct 31, 2012, at 7:37 AM, Alexandre Fouche
> <[EMAIL PROTECTED]
> <mailto:[EMAIL PROTECTED]>> wrote:
>
>> These instances have no swap. I tried 5 or 6 times in a row, and
>> modified the yarn.nodemanager.resource.memory-mb but it did not help.
>> Later on, i'll replace the openjdk with the Oracle java SE 1.6.31 to
>> see if it improves overall performance.

How many RAM do you have, and how much of it is assigned to your Hadoop
services?

>> Now i am running everything on medium instances for prototyping, and
>> while this is better, i still find it abusively slow. Maybe bad
>> hadoop performance on less than xlarge memory instances is to be
>> expected on EC2 ?
Are you using Hadoop on top of EC2 or are you using the EMR service?

>>
>>
>> --
>> Alexandre Fouche
>> Lead operations engineer, cloud architect
>> http://www.cleverscale.com | @cleverscale
>> Sent with Sparrow <http://www.sparrowmailapp.com/?sig>
>>
>> On Monday 29 October 2012 at 20:04, Michael Segel wrote:
>>
>>> how many times did you test it?
>>>
>>> need to rule out aberrations.
>>>
>>> On Oct 29, 2012, at 11:30 AM, Harsh J <[EMAIL PROTECTED]
>>> <mailto:[EMAIL PROTECTED]>> wrote:
>>>
>>>> On your second low-memory NM instance, did you ensure to lower the
>>>> yarn.nodemanager.resource.memory-mb property specifically to avoid
>>>> swapping due to excessive resource grants? The default offered is 8 GB
>>>> (>> 1.7 GB you have).
>>>>
>>>> On Mon, Oct 29, 2012 at 8:42 PM, Alexandre Fouche
>>>> <[EMAIL PROTECTED]
>>>> <mailto:[EMAIL PROTECTED]>> wrote:
>>>>> Hi,
>>>>>
>>>>> Can someone give some insight on why a "distcp" of 600 files of a few
>>>>> hundred bytes from s3n:// to local hdfs is taking 46s when using a
>>>>> yarn-nodemanager EC2 instance with 16GB memory (which by the way i
>>>>> think is
>>>>> jokingly long), and taking 3mn30s when adding a second
>>>>> yarn-nodemanager (a
>>>>> small instance with 1.7GB memory) ?
>>>>> I would have expected it to be a bit faster, not 5xlonger !
>>>>>
>>>>> I have the same issue when i stop the small instance nodemanager
>>>>> and restart
>>>>> it to join the processing after the big nodemanager instance was
>>>>> already
>>>>> submitted the job.
>>>>>
>>>>> I am using Cloudera latest Yarn+HDFS on Amazon (rebranded Centos 6)
>>>>>
>>>>> #Staging 14:58:04 root@datanode2:hadoop-yarn: rpm -qa |grep hadoop
>>>>> hadoop-hdfs-datanode-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>> hadoop-mapreduce-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>> hadoop-0.20-mapreduce-0.20.2+1261-1.cdh4.1.1.p0.4.el6.x86_64
>>>>> hadoop-yarn-nodemanager-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>> hadoop-mapreduce-historyserver-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>> hadoop-hdfs-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>> hadoop-client-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>> hadoop-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>> hadoop-yarn-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
>>>>>
>>>>>
>>>>> #Staging 14:39:51 root@resourcemanager:hadoop-yarn:
>>>>> HADOOP_MAPRED_HOME=/usr/lib/hadoop-mapreduce time hadoop distcp
>>>>> -overwrite
>>>>> s3n://xxx:[EMAIL PROTECTED]ev <mailto:[EMAIL PROTECTED]ev>/*
>>>>> hdfs:///tmp/something/a
>>>>>
>>>>> 12/10/29 14:40:12 INFO tools.DistCp: Input Options:
>>>>> DistCpOptions{atomicCommit=false, syncFolder=false,
>>>>> deleteMissing=false,
>>>>> ignoreFailures=false, maxMaps=20, sslConfigurationFile='null',
>>>>> copyStrategy='uniformsize', sourceFileListing=null,
>>>>> sourcePaths=[s3n://xxx:[EMAIL PROTECTED]ev
>>>>> <mailto:[EMAIL PROTECTED]ev>/*],
>>>>> targetPath=hdfs:/tmp/something/a}
>>>>> 12/10/29 14:40:18 WARN conf.Configuration: io.sort.mb is deprecated.
>>>>> Instead, use mapreduce.task.io.sort.mb
Marcos Luis Ortï¿½z Valmaseda
about.me/marcosortiz <http://about.me/marcosortiz>
@marcosluis2186 <http://twitter.com/marcosluis2186>

10mo. ANIVERSARIO DE LA CREACION DE LA UNIVERSIDAD DE LAS CIENCIAS INFORMATICAS...
CONECTADOS AL FUTURO, CONECTADOS A LA REVOLUCION

http://www.uci.cu
http://www.facebook.com/universidad.uci
http://www.flickr.com/photos/universidad_uci

Re: Insight on why distcp becomes slower when adding nodemanager

Alexandre Fouche 2012-11-01, 20:01

I was using Yarn and HDFS on EC2 and EBS, with default memory settings.

I have just read in the Hadoop guide a list of hadoop benchmark jars. I guess i'll use Whirr to create a canned hadoop cluster on ec2, and run these benchmarks. So i will have a baseline to which i can compare. Then i'll compare with my own install of hadoop stack
--
Alexandre Fouche
Lead operations engineer, cloud architect
http://www.cleverscale.com | @cleverscale
Sent with Sparrow (http://www.sparrowmailapp.com/?sig)
On Wednesday 31 October 2012 at 21:27, Marcos Ortiz wrote:

>
> On 10/31/2012 02:23 PM, Michael Segel wrote:
> > Not sure.
> >
> > Lots of things can effect your throughput.
> > Networking is my first guess. Which is why I asked about the number of times you've run the same test to see if there is a wide variation in timings.
> >
> > On Oct 31, 2012, at 7:37 AM, Alexandre Fouche <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
> > > These instances have no swap. I tried 5 or 6 times in a row, and modified the yarn.nodemanager.resource.memory-mb but it did not help. Later on, i'll replace the openjdk with the Oracle java SE 1.6.31 to see if it improves overall performance.
> How many RAM do you have, and how much of it is assigned to your Hadoop services?
>
> > > Now i am running everything on medium instances for prototyping, and while this is better, i still find it abusively slow. Maybe bad hadoop performance on less than xlarge memory instances is to be expected on EC2 ? Are you using Hadoop on top of EC2 or are you using the EMR service?
>
> > >
> > >
> > > --
> > > Alexandre Fouche
> > > Lead operations engineer, cloud architect
> > > http://www.cleverscale.com | @cleverscale
> > > Sent with Sparrow (http://www.sparrowmailapp.com/?sig)
> > >
> > >
> > > On Monday 29 October 2012 at 20:04, Michael Segel wrote:
> > >
> > > > how many times did you test it?
> > > >
> > > > need to rule out aberrations.
> > > >
> > > > On Oct 29, 2012, at 11:30 AM, Harsh J <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
> > > >
> > > > > On your second low-memory NM instance, did you ensure to lower the
> > > > > yarn.nodemanager.resource.memory-mb property specifically to avoid
> > > > > swapping due to excessive resource grants? The default offered is 8 GB
> > > > > (>> 1.7 GB you have).
> > > > >
> > > > > On Mon, Oct 29, 2012 at 8:42 PM, Alexandre Fouche
> > > > > <[EMAIL PROTECTED] (mailto:[EMAIL PROTECTED])> wrote:
> > > > > > Hi,
> > > > > >
> > > > > > Can someone give some insight on why a "distcp" of 600 files of a few
> > > > > > hundred bytes from s3n:// to local hdfs is taking 46s when using a
> > > > > > yarn-nodemanager EC2 instance with 16GB memory (which by the way i think is
> > > > > > jokingly long), and taking 3mn30s when adding a second yarn-nodemanager (a
> > > > > > small instance with 1.7GB memory) ?
> > > > > > I would have expected it to be a bit faster, not 5xlonger !
> > > > > >
> > > > > > I have the same issue when i stop the small instance nodemanager and restart
> > > > > > it to join the processing after the big nodemanager instance was already
> > > > > > submitted the job.
> > > > > >
> > > > > > I am using Cloudera latest Yarn+HDFS on Amazon (rebranded Centos 6)
> > > > > >
> > > > > > #Staging 14:58:04 root@datanode2:hadoop-yarn: rpm -qa |grep hadoop
> > > > > > hadoop-hdfs-datanode-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > > > > hadoop-mapreduce-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > > > > hadoop-0.20-mapreduce-0.20.2+1261-1.cdh4.1.1.p0.4.el6.x86_64
> > > > > > hadoop-yarn-nodemanager-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > > > > hadoop-mapreduce-historyserver-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > > > > hadoop-hdfs-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > > > > hadoop-client-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
> > > > > > hadoop-2.0.0+545-1.cdh4.1.1.p0.5.el6.x86_64
(http://www.uci.cu/)>
(http://www.uci.cu/)