Yarn Performance: submitting applications slow down application scheduling?

[bo yang <bo...@gmail.com>]

Hi,

I am doing some performance test on YARN to see its performance on large
scale application submission and scheduling. I find the application
scheduling slowed down when I continuously submitted jobs in a high speed
(e.g. 20 applications per second).

Anyone has any experience with this? Is there any way for tuning up Yarn
for large scale applications (e.g. 20 applications per second for a whole
day)?


A bit more details about my test...

I am using capacity scheduler. Also I use my custom container executor to
run the application within same process of node manager (not creating new
process). In my test, I set a dummy application which runs and finishes
immediately. I use one resource manager node and five node manager nodes.

When there are 10,000 pending applications in Resource Manager, the
application completion speed is 1-3 per second when I still keep submitting
applications. If I stop submitting new applications, the application
completion speed is around 10 per second. So I feel the application
submission slows down the overall application dispatching.

Thanks,
Bo