When I have seen that sawtooth pattern before, it was because the cache hit 
rate was low. Items ejected from the cache become garbage. If that happens at a 
fairly constant rate, you get that slope of increasing memory usage.

wunder

On Oct 7, 2011, at 8:48 AM, Tristan Roddis wrote:

> We have a Solr server running on 64-bit Windows 2008, and under the default 
> Tomcat settings it seems to be using a small fraction of the available memory 
> (fluctuating between 100Mb and 300Mb: see trace from VisualVM at 
> http://imgur.com/4Om2El&tqLKW <http://imgur.com/4Om2El&tqLKW> )
> 
> I was advised that, to increase performance, we should increase the amount of 
> memory allocated to Tomcat so that Solr could use this to cache query results 
> and so reduce CPU usage and disk I/O. So, I set both the min and max RAM 
> allocation for Tomcat to 10Gb, which is well within the amount allocated to 
> the server (16Gb). I tried this on another machine, set a simulated load of 
> lots of concurrent users running on our application, and the VisualVM results 
> can be seen at http://imgur.com/4Om2E&tqLKWl <http://imgur.com/4Om2E&tqLKWl>
> 
> These results were very different: in the first half of the graph, when the 
> system was under load, the amount of allocated heap size was fixed at 10Gb, 
> which I would expect, but the amount of used heap climbed gradually to around 
> 3Gb, but then shot right down again to something close to zero every minute 
> or so, presumably due to the GC kicking in, giving the sawtooth pattern you 
> can see in the screenshot above.
> 
> My question is: is this normal and expected? I would have thought that the 
> amount of used heap would climb to somewhere much nearer the maximum 
> available of 10Gb, rather than peaking at a mere 3Gb. And also that any 
> garbage collection would only deallocate part of the memory rather than 
> knocking it right back to zero.
> 
> Essentially, why isn't Solr using more of the memory assigned, and using it 
> constantly rather than continually clearing the cache?
> 
> Thanks,
> 
> -Tristan.




Reply via email to