> I have personally found that increasing the size of the young generation 
> (Eden) is beneficial to Solr,

I've seen the same thing - I think it's because requests create a lot
of short lived objects and if the eden is not large enough, a lot of
those objects will make it to the tenured space, which is basically an
alg fail.

It's not a bad knob to tweak, because if you just keep raising the
heap, you can wastefully keep giving more unnecessary RAM to the
tenured space when you might only want to give more to the eden space.

- Mark


On Wed, Nov 21, 2012 at 11:00 AM, Shawn Heisey <s...@elyograg.org> wrote:
> On 11/21/2012 12:36 AM, stockii wrote:
>>
>> okay. i will try out more RAM.
>>
>> i am using not much caching because of "near-realt-time"-search. in this
>> case its better to increase xmn or only xmx and xms?
>
>
> I have personally found that increasing the size of the young generation
> (Eden) is beneficial to Solr, at least if you are using the parallel GC
> options.  I theorize that the collector for the young generation is more
> efficient than the full GC, but that's just a guess.  When I started doing
> that, the amount of time my Solr JVM spent doing garbage collection went way
> down, even though the number of garbage collections went up.
>
> Lately I have been increasing the Eden size by using -XX:NewRatio=1 rather
> than an explicit value on -Xmn.  This has one advantage - if you change the
> min/max heap size, the same value for NewRatio will still work.
>
> Here are the options that I am currently using in production with Java6:
>
> -Xms4096M
> -Xmx8192M
> -XX:NewRatio=1
> -XX:+UseParNewGC
> -XX:+UseConcMarkSweepGC
> -XX:+CMSParallelRemarkEnabled
>
> Here is what I am planning for the future with Solr4 and beyond with Java7,
> including an environment variable for Xmx. Due to the experimental nature of
> the G1 collector, I would only trust it with the latest Java releases,
> especially for Java6.  The Unlock option is not required on Java7, only
> Java6.
>
> -Xms256M
> -Xmx${JMEM}
> -XX:+UnlockExperimentalVMOptions
> -XX:+UseG1GC
>
> Thanks,
> Shawn
>



-- 
- Mark

Reply via email to