With the last lot of changes to WeldClassImpl memory usage is down by about 60Mb ( ~30%)
for the 5000 class benchmark. The fields that I have removed are only used at startup of
during de-serialisation.
The effect on startup time is not noticeable, and I think that cost of de serialisation
will be significantly higher than any overhead imposed by these changes, although I have
not benchmarked this yet.
There is still plenty more that can be done here, I think reducing the memory usage by 50%
should be easily achievable.
Stuart
On 10/11/2010, at 9:57 PM, David Allen wrote:
Stuart,
Excellent! There shouldn't be too many hash maps left in WeldClassImpl. Most of
these have already been replaced with array set implementations; the ones left are a bit
more complicated to replace due to use cases.
Pete also wanted to redo the reflection data that we maintain and cache. It should be
possible to release most of this cached data after bootstrapping the app. As it is now,
everything is kept in memory permanently with the app which really isn't necessary.
I'm supposed to be working on this, but I don't think I'll have much time
till January 2011 to really do much. Keep us posted and I'll try to take a look on
weekends or whenever I have a little time.
Thanks,
David
On Wed, Nov 10, 2010 at 11:12 AM, Stuart Douglas <stuart.w.douglas(a)gmail.com>
wrote:
Clearing the resolvers after startup saves about 42Mb in the 5000 bean case (~20% of the
memory attributable to weld, and ~10% of total memory usage).
Stuart
On Wed, Nov 10, 2010 at 8:12 PM, Stuart Douglas <stuart.w.douglas(a)gmail.com>
wrote:
I just ran some very quick and dirty profiling with the latest Jbossas and the results
are as follows:
Beans Startup Time Startup (WELDX) Memory Usage Mem Usage(no beans.xml)
No Deployment 17 135
20 20 22 149
500 24 26 178
2000 35 43 265
5000 87 104 440 210
So jboss uses 135Mb normally, and 210Mb when a war with 5000 classes is deployed that
does not have beans.xml. When you add weld to the mix the memory usage jumps by 230Mb to
440Mb.
According to yjp WeldClassImpl (and it's retained WeldMethod/Field etc) is
responsible for 120Mb of this. Other major culprits seem to be TypeSafeObserverResolver at
24Mb (as it is caching ProcessAnnotatedType<Bean*> * 5000) and
TypeSafeDecoratorResolver at 13Mb. Not much else stands out.
The beans used where quite simple (1 injection point, 7 fields, 6 methods), no normal
scoped beans, no interceptors, not decorators. Weldx does have a notable effect on startup
time, which I will also investigate.
I don't think it will be to hard to significantly reduce this. Reducing the number of
HashMap's in WeldClassImpl (and replacing some with ImmutableArraySet) should give a
significant gain, and clearing the TypeSafeObserverResolver and TypeSafeDecoratorResolver
after startup should also save around 40Mb. I'll try and do some work this week and
see how much I can get this down.
Stuart
On 10/11/2010, at 8:48 AM, Pete Muir wrote:
> I'm about to post a blog about this.
>
> On 9 Nov 2010, at 21:43, Lincoln Baxter, III wrote:
>
>> Are these points valid?
>>
>> If so, are we aware of them? Just trying to raise awareness to what people are
saying out in the world. I have noticed a relatively high memory footprint in Seam Forge,
using Weld SE.
>>
>>
http://www.dzone.com/links/r/cdi_a_major_risk_factor_in_java_ee_6.html
>>
>> Is there anything we can address here and attempt to demystify this blog?
>>
>> --
>> Lincoln Baxter, III
>>
http://ocpsoft.com
>>
http://scrumshark.com
>> "Keep it Simple"
>> _______________________________________________
>> weld-dev mailing list
>> weld-dev(a)lists.jboss.org
>>
https://lists.jboss.org/mailman/listinfo/weld-dev
>
>
> _______________________________________________
> weld-dev mailing list
> weld-dev(a)lists.jboss.org
>
https://lists.jboss.org/mailman/listinfo/weld-dev
_______________________________________________
weld-dev mailing list
weld-dev(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/weld-dev