Hi Dmitry,
Seems to be an interesting case, would like some more clarifications in this 
regard :
1. How many NM's ? Is it a hetergenous cluster or all the nodes have same 
resource capacity ? by 3000 cores if same config then expecting around 100 
nodes, am i correct ?
2. How many applications are running and how many have got finished (basically 
available in RM) ? By 35000 you mean finished and running applications ?
3. Weather after some time, tasks are getting assigned ? Also is it only this 
host not getting assigned or no other host also gets any containers assigned ?

I suspect this issue might be similar to YARN-3990, hence the above questions. 
Further you can check the RM logs and inform weather you see some similar logs 
as below

2015-07-29 19:39:03,416 | INFO  | AsyncDispatcher event handler | Size of 
event-queue is 14000 | AsyncDispatcher.java:235
2015-07-29 19:39:03,417 | INFO  | AsyncDispatcher event handler | Size of 
event-queue is 15000 | AsyncDispatcher.java:235

Regards,
+ Naga


________________________________
From: Dmitry Sivachenko [[email protected]]
Sent: Wednesday, September 23, 2015 03:57
To: [email protected]
Subject: node remains unused after reboot

Hello!

I am using hadoop-2.7.1. I have a large map job running (total cores available 
on the cluster about 3000, total tasks 35000).
In the middle of this process one server reboots.

After reboot, nodemanager starts successfully end registers with resource 
manager:
2015-09-23 01:06:24,656 INFO  [main] nodemanager.NodeStatusUpdaterImpl 
(NodeStatusUpdaterImpl.java:registerWithRM(311)) - Notifying ContainerManager 
to unblock new container-requests

In YARN web-interface I see this host as active, but VCores used remains zero 
(see screenshot).
But the map job mentioned is still running and have about 12000 pending tasks.

Why this host does not receive tasks to run?

PS: I recently upgraded from 2.4.1 and I did not notice such a problem with 
2.4.1: new tasks were spawning immediately after reboot.

Thanks!




[cid:[email protected]]

Reply via email to