hadoop-hdfs-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Brahma Reddy Battula <brahmareddy.batt...@huawei.com>
Subject RE: Max Parallel task executors
Date Mon, 09 Nov 2015 12:18:10 GMT

I'm glad to hear it helped.

Thanks & Regards

 Brahma Reddy Battula

From: sandeep das [yarnhadoop@gmail.com]
Sent: Monday, November 09, 2015 11:54 AM
To: user@hadoop.apache.org
Subject: Re: Max Parallel task executors

After increasing yarn.nodemanager.resource.memory-mb to 24 GB more number of parallel map
tasks are being spawned. Its resolved now.
Thanks a lot for your input.


On Mon, Nov 9, 2015 at 9:49 AM, sandeep das <yarnhadoop@gmail.com<mailto:yarnhadoop@gmail.com>>
BTW Laxman according to the formula that you had provided it turns out that only 8 jobs per
node will be initiated which is matching with what i'm seeing on my setup.

min (yarn.nodemanager.resource.memory-mb / mapreduce.[map|reduce].memory.mb,
     yarn.nodemanager.resource.cpu-vcores / mapreduce.[map|reduce].cpu.vcores)

yarn.nodemanager.resource.memory-mb: 16 GB

mapreduce.map.memory.mb: 2 GB

yarn.nodemanager.resource.cpu-vcores: 80

mapreduce.map.cpu.vcores: 1

So if apply the formula then min(16/2, 80/1) -> min(8,80) -> 8

Should i reduce memory per map operation or increase memory for resource manager?

On Mon, Nov 9, 2015 at 9:43 AM, sandeep das <yarnhadoop@gmail.com<mailto:yarnhadoop@gmail.com>>
Thanks Brahma and Laxman for your valuable input.

Following are the statistics available on YARN RM GUI.

Memory Used : 0 GB
Memory Total : 64 GB (16*4 = 64 GB)
VCores Used: 0
VCores Total: 320 (Earlier I had mentioned that I've configured 40 Vcores but recently I increased
to 80 that's why its appearing 80*4 = 321)

Note: These statistics were captured when there was no job running in background.

Let me know whether it was sufficient to nail the issue. If more information is required please
let me know.


On Fri, Nov 6, 2015 at 7:04 PM, Brahma Reddy Battula <brahmareddy.battula@huawei.com<mailto:brahmareddy.battula@huawei.com>>

The formula for determining the number of concurrently running tasks per node is:

min (yarn.nodemanager.resource.memory-mb / mapreduce.[map|reduce].memory.mb,
     yarn.nodemanager.resource.cpu-vcores / mapreduce.[map|reduce].cpu.vcores) .

For you scenario :

As you told yarn.nodemanager.resource.memory-mb is configured to 16 GB and yarn.nodemanager.resource.cpu-vcores
configured to 40. and I am thinking
mapreduce.map/reduce.memory.mb, mapreduce.map/reduce.cpu.vcores default values.

min (16GB/1GB,40Core/1Core )=16 tasks for Node. Then total should be 16*4=64  (63+1AM)..

I am thinking, Two Nodemanger's are unhealthy (OR) you might have configured mapreduce.map/reduce.memory.mb=2GB(or
5 core).

As laxman pointed you can post RMUI or you can cross check like above.

Hope this helps.

Thanks & Regards

 Brahma Reddy Battula

From: Laxman Ch [laxman.lux@gmail.com<mailto:laxman.lux@gmail.com>]
Sent: Friday, November 06, 2015 6:31 PM
To: user@hadoop.apache.org<mailto:user@hadoop.apache.org>
Subject: Re: Max Parallel task executors

Can you please copy paste the cluster metrics from RM dashboard.
Its under http://rmhost:port/cluster/cluster

In this page, check under Memory Total vs Memory Used and VCores Total vs VCores Used

On 6 November 2015 at 18:21, sandeep das <yarnhadoop@gmail.com<mailto:yarnhadoop@gmail.com>>
HI Laxman,

Thanks for your response. I had already configured a very high value for yarn.nodemanager.resource.cpu-vcores
e.g. 40 but still its not increasing more number of parallel tasks to execute but if this
value is reduced then it runs less number of parallel tasks.

As of now yarn.nodemanager.resource.memory-mb is configured to 16 GB and yarn.nodemanager.resource.cpu-vcores
configured to 40.

Still its not spawning more tasks than 31.

Let me know if more information is required to debug it. I believe there is upper limit after
which yarn stops spawning tasks. I may be wrong here.


On Fri, Nov 6, 2015 at 6:15 PM, Laxman Ch <laxman.lux@gmail.com<mailto:laxman.lux@gmail.com>>
Hi Sandeep,

Please configure the following items to the cores and memory per node you wanted to allocate
for Yarn containers.
Their defaults are 8 cores and 8GB. So that's the reason you were stuck at 31 (4nodes * 8cores
- 1 AppMaster)


On 6 November 2015 at 17:59, sandeep das <yarnhadoop@gmail.com<mailto:yarnhadoop@gmail.com>>
May be to naive to ask but How do I check that?
Sometimes there are almost 200 map tasks pending to run but at a time only 31 runs.

On Fri, Nov 6, 2015 at 5:57 PM, Chris Mawata <chris.mawata@gmail.com<mailto:chris.mawata@gmail.com>>

Also check that you have more than 31 blocks to process.

On Nov 6, 2015 6:54 AM, "sandeep das" <yarnhadoop@gmail.com<mailto:yarnhadoop@gmail.com>>
Hi Varun,

I tried to increase this parameter but it did not increase number of parallel tasks but if
It is decreased then YARN reduces number of parallel tasks. I'm bit puzzled why its not increasing
more than 31 tasks even after its value is increased.

Is there any other configuration as well which controls on how many maximum tasks can execute
in parallel?


On Tue, Nov 3, 2015 at 7:29 PM, Varun Vasudev <vvasudev@apache.org<mailto:vvasudev@apache.org>>
The number of parallel tasks that are run depends on the amount of memory and vcores on your
machines and the amount of memory and vcores required by your mappers and reducers. The amount
of memory can be set via yarn.nodemanager.resource.memory-mb(the default is 8G). The amount
of vcores can be set via yarn.nodemanager.resource.cpu-vcores(the default is 8 vcores).


From: sandeep das <yarnhadoop@gmail.com<mailto:yarnhadoop@gmail.com>>
Reply-To: <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Date: Monday, November 2, 2015 at 3:56 PM
To: <user@hadoop.apache.org<mailto:user@hadoop.apache.org>>
Subject: Max Parallel task executors

Hi Team,

I've a cloudera cluster of 4 nodes. Whenever i submit a job my only 31 parallel tasks are
executed whereas my machines have more CPU available but still YARN/AM does not create more

Is there any configuration which I can change to start more MAP/REDUCER task in parallel?

Each machine in my cluster has 24 CPUs.




View raw message