[
https://issues.apache.org/jira/browse/YARN-7672?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16298689#comment-16298689
]
Wei Yan commented on YARN-7672:
-------------------------------
bq. All of NM and AM simulators are all cpu type of task. So cpu.load will go
up to 100+ (only 32 cores) And as we know, Scheduler will also use one process
for allocating resources.
So in that case, even separating into two hosts, the host running NM/AM
simulators still hit the CPU bottleneck, right? Although the Scheduler doesn't
need to compete with simulators.
Another interesting idea is to launch a large MapReduce job (like 5000
containers), each container runs as NM/AM simulator, to issue requests to the
real RM. similar to the idea for HDFS [Dynamometer |
https://lists.apache.org/thread.html/7223d22fbc26e055369695f83395e9a7767043f7245af25df385b535@%3Chdfs-dev.hadoop.apache.org%3E].
But this involves more complext setup..
> hadoop-sls can not simulate huge scale of YARN
> ----------------------------------------------
>
> Key: YARN-7672
> URL: https://issues.apache.org/jira/browse/YARN-7672
> Project: Hadoop YARN
> Issue Type: Improvement
> Reporter: zhangshilong
> Assignee: zhangshilong
> Attachments: YARN-7672.patch
>
>
> Our YARN cluster scale to nearly 10 thousands nodes. We need to do scheduler
> pressure test.
> Using SLS,we start 2000+ threads to simulate NM and AM. But cpu.load very
> high to 100+. I thought that will affect performance evaluation of
> scheduler.
> So I thought to separate the scheduler from the simulator.
> I start a real RM. Then SLS will register nodes to RM,And submit apps to RM
> using RM RPC.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
