Well, what I have been trying to is send a "Cluster Pressure Event" to the 
Spark Applications running on a cluster to allow for dynamic re-sizing of Spark 
applications running on the cluster. It is a 4th year project that I am working 
on for my university program. We have the Spark side implemented where spark 
applications are set to have a fair share value for the number of executors 
along with the already implemented max and min values. This way, a spark app 
will grow until it has it's max and then throttle back to it's fair share if a 
Cluster Pressure Event is triggered. The motivation behind this, is to allow 
for interactive jobs to come onto the cluster - so that they don't have to wait 
for the other jobs to complete. Currently the only way for a spark application 
to "shrink down" is if the executors time out for inactivity. We have defined 
cluster pressure as (pending resources + allocated resources / total resources 
> 1) meaning that there has been more requests for resources than what is 
available - so we need to release some. 
I had attempted previously to do this inside the resource manager - but I could 
not figure out how to get events to the client from the RM. So now I am trying 
to do this client side, but I don't have access to this data (all I need is 
available memory, pending memory, and total memory available - or something 
similar).
If you could provide some insight it would be greatly appreciated. I have been 
having trouble with this for months now.
Thanks,Charles.

> From: [email protected]
> To: [email protected]
> CC: [email protected]
> Subject: Re: Question about QueueMetrics/Cluster Metrics
> Date: Sat, 12 Mar 2016 12:05:52 -0800
> 
> On 3/12/16 12:00 PM, Charlie Wright wrote:
> > Is there a way to get the root QueueMetrics on the client side?-Charles.    
> >                                 
> 
> The QueueMetrics objects aren't exposed outside of the RM.  What are you 
> trying to accomplish?  If you just want to access the metrics data, you 
> can configure a RollingFileSystemSink to write the metrics into HDFS so 
> you can pick them up.
> 
> Daniel
                                          

Reply via email to