[osg-users] Slow optimization and OpenFlight

[Andreas Ekstrand]

Forgot to mention that the geometry before and after optimization is 
identical in the different OSG versions, so the optimizer accomplishes 
the same thing, it just takes about 10 times longer.
/Andreas


On 2017-11-24 14:10, Andreas Ekstrand wrote:
> Hi,
>
> Some more results from my investigations:
>
> Back in February, the MergeGeometryVisitor was changed to be applied 
> on Group instead of Geode by Jannik Heller (in CC). This has resulted 
> in substantially longer optimization time of my loaded OpenFlight 
> files with a large number of polygons. I assume it's a general 
> performance problem with many geodes/geometries.
>
> The removeChild call from MergeGeometryVisitor ::mergeGroup is the 
> main culprit and takes about 87% of the time:
> group.removeChild(*ditr);
>
> Does Robert or Jannik have any idea about this? I don't feel 
> comfortable enough with the code to make any changes, but this 
> prevents me from upgrading my software to the latest OSG version, 
> while I can't go back either since I need other fixes in 3.5.8. So I'm 
> a bit stuck and would appreciate any help!
>
> Regards,
> Andreas
>
>
> On 2017-11-19 17:13, Andreas Ekstrand wrote:
>> Hi Robert,
>>
>> Yes, the model is ineffective in the sense that it has 150 000 
>> separate triangles on the same level in the scene graph, that's the 
>> nature of basic usage of OpenFlight and I guess that's why the plugin 
>> applies an optimization of its own. But this could be optimized and 
>> merged much faster before and I'm just looking for a way to get it to 
>> behave like that again, to keep my software from being slower in a 
>> new version.
>>
>> Some results from VTune, where I have manually filtered and presented 
>> the top two culprits in 3.5.1 compared to 3.5.8 (where I gave up 
>> after 2 minutes):
>>
>> 3.5.8: TOTAL: 121.4s
>>
>>   * Registry::read -> Optimizer::optimize ->
>>     MergeGeometryVisitor::mergeGroup -> Group::removeChildren ->
>>     OpenThreads::Atomic::operator++ (78.5s)
>>   * Registry::read -> Optimizer::optimize ->
>>     MergeGeometryVisitor::mergeGroup -> Group::removeChildren ->
>>     OpenThreads::Atomic::operator-- (27.4s)
>>
>> 3.5.1: TOTAL: 12.7s
>>
>>   * Registry::read -> Optimizer::optimize ->
>>     Optimizer::StateVisitor::optimize -> Node::setStateSet ->
>>     StateSet::~StateSet -> StateSet::clear ->
>>     StateAttribute::removeParent (2.4s)
>>   * Registry::read -> Optimizer::optimize ->
>>     Optimizer::MergeGeodesVisitor::mergeGeodes -> Node::~Node ->
>>     Node::setStateSet -> StateAttribute::removeParent (2.2s)
>>
>> The MergeGeometryVisitor was applied in 3.5.1 as well but didn't take 
>> more than about 0.06s. I guess the 3.5.1 results are generally less 
>> interesting but at least they show that MergeGeometryVisitor was not 
>> a problem there.
>>
>> I don't know if this gives you some ideas directly or if we need to 
>> keep digging. I will have another look shortly either way, but of 
>> course hoping for you or someone else to come up with some bright 
>> ideas to shorten the time to fixing this problem.
>>
>> Regards,
>> Andreas
>>
>>
>> On 2017-11-19 16:28, Robert Osfield wrote:
>>> HI Andreas,
>>>
>>> I haven't had a chance to dig further.
>>>
>>> One curious thing I noticed is that when I enabled verbose debug
>>> output there was lots of buffer objects being created and destroyed
>>> during the optimisation step.  osg::Drawable now assigns
>>> VetextArrayState and VBO's by default for osg::Geometry - this is
>>> required to make VAO support possible.  The large number of these
>>> operations suggest lots of creation and merging of osg::Geometry for
>>> this dataset.  There is chance tthat this change alone may be causing
>>> a performance slow down.
>>>
>>> However, I don't think creation/deletion of osg::Geometry is the crux
>>> of the problem, and may not even be the optimization step - it could
>>> well be down to the nature of the scene graph being created by the
>>> OpenFlight plugin for this dataset.  I suspect the source data is
>>> stored in a way that is really inefficient to handle for real-time.
>>> It might be that the OpenFlight plugin just isn't handling the data
>>> well.
>>>
>>> As a general rule, running the Optimizer is a last resort for fixing
>>> really bad datasets, ideally datasets should be created in a form that
>>> is appropriate for decent performance right from the start.
>>>
>>> Robert.
>>> _______________________________________________
>>> osg-users mailing list
>>> osg-users at lists.openscenegraph.org
>>> http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
>>>
>>
>>
>>
>> _______________________________________________
>> osg-users mailing list
>> osg-users at lists.openscenegraph.org
>> http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
>
>
>
> _______________________________________________
> osg-users mailing list
> osg-users at lists.openscenegraph.org
> http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openscenegraph.org/pipermail/osg-users-openscenegraph.org/attachments/20171124/99c79b5a/attachment.html>