Hi, JP,
On 01/26/2012 01:53 AM, J.P. Delport wrote:
very interesting and thorough investigation. I can't really add anything and unfortunately don't have representative hardware to test on. It would be interesting to see if different bus configs (motherboards/chipsets) make any difference.
The target hardware is a dual Xeon system with a 5400 chipset (the last chipset with a FSB) and a pair of QuadroPlex D2's. I've been working on an Alienware desktop with an X58 chipset and two GeForce 260's. One would expect that the X58 with its QPI interconnect would succumb to bus contention later than the 5400 (the specs say so, at least), but on the largest data set, it appears that the X58 is still having the scaling problem, while on the workstation, the bottleneck has moved elsewhere.
I also would not rule out the NVidia driver doing silly things. I remember once looking at some of the glue code for the binary blob and seeing TLB flushes that caused a Xenomai kernel to fail ito timing in interesting ways. Have you tried OProfile to get some idea of time spent in the driver?
I'm not familiar with OProfile. I'll take a look. Thanks for the reply! --"J" _______________________________________________ osg-users mailing list osg-users@lists.openscenegraph.org http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org