We are trying to optimize very short GC latencies (hopefully around 10ms... 
every minute or hopefully more ... and the best way to do it - is not create so 
many objects). The use case is low throughput - low latency, which is not the 
usual use case I guess. We are still experimenting, most of the objects are 
storm related.



________________________________
From: Bobby Evans <[email protected]>
Sent: Wednesday, November 19, 2014 10:37 PM
To: [email protected]
Subject: Re: Minimizing stats.clj garbage collection effects

Only topology.stats.sample.rate might have an impact on it.  The other config 
has nothing to do with stats.clj.
>From reading through the code it looks like the RollingWindow record being 
>used to keep track of stats is creating new objects each time it is updated.  
>This is a very functional way of doing it, but I am not sure it is best for 
>GC.  Is GC causing you a lot of issues?

- Bobby


On Wednesday, November 19, 2014 8:40 AM, Itai Frenkel <[email protected]> wrote:


Hi,

We've enabled yourkit profiler garbage collection statistics in order to reduce 
GC latencies. It seems that stats.clj contribute a big part of the GCed 
objects. What settings should I tweak in order to reduce stats object creation 
? How would it affect the ui reporting accuracy ?

topology.stats.sample.rate: 0.05
topology.builtin.metrics.bucket.size.secs: 60

Regards,
Itai




Reply via email to