On 01/27/2012 02:12 PM, Paul Martz wrote:
Sorry, I was going off your post that read "This leads me to believe that bus
contention was causing the lack of scalability." I was just trying to outline a
way to validate that suspicion. If you've already ruled out bus contention, then
disregard my post.
Hi, Paul,
I'm not dismissing your post at all. It's certainly worth thinking
about the overall bandwidth, and how much of it might be consumed by the
GL command stream. I like your equation as a general guideline. One
just has to keep in mind that 500MB/sec is the bandwidth for a single
PCI-e 2.0 lane. Most graphics cards get 8 lanes, if not 16, and that's
per card. You also have to account for the CPU interconnect bandwidth,
which may put an upper limit on how much data you can transfer with
multiple cards (theoretically 12.8GB/s on this hardware).
In this particular case, I think Tim is right. There's not enough data
in the command stream to get near the bus bandwidth. In this case,
there are two 8x PCI-e cards (total of 8GB/s bandwidth) feeding 4 GPU's
(two QuadroPlex D2 units). Using your figures:
2 * 8 * 500 MB/s
----------------------------- = 5648 Hz
4 * ( (5500 * 64 B) + 2048 B)
So, unless I'm missing something, the problem must be elsewhere :-)
--"J"
_______________________________________________
osg-users mailing list
[email protected]
http://lists.openscenegraph.org/listinfo.cgi/osg-users-openscenegraph.org
