Hi Bob, Last summer when testing multiple GPU rendering I found that if I forced the serialization of the draw dispatch I actually got better performance. This suggest a performance problem in the driver, or it could also have been a hardware contention. Either way even on my quad core, dual GPU machine serializing draw dispatch was a winner - against what I would have expected for a properly balanced system.
I did call for testing of draw serialization from the community, but alas got 0 feedback on this issue, so had to go with the only data point that I had - my own experience with my system, and choice the default to serialize the draw. This option is configurable via the osg::DisplaySetting classes setSerializeDrawDispatch method, or its default setting via the env var OSG_SERIALIZE_DRAW_DISPATCH, if you set this env var to OFF, then it'll allow the draw dispatch to run multi-threaded. For your tests it's unlikely to make too much difference as draw dispatch for a textured cube will be far less than 1ms, so even doing for draw traversals in series won't be something that breaks frame. I'll do some more testing today to see if things have changed in the last year, feedback on other system would be really really useful. Robert. _______________________________________________ osg-users mailing list [email protected] http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
