[eq-dev] AFR-style DPlex on a Dual-GPU machine

[Stefan Eilemann]

On 5. Jan 2009, at 11:08, Mathias Heyer wrote:

> First I tried a SFR-style config with a classic 2D load balancer. It  
> worked almost right out of the box. It didn’t give a performance  
> boost, though. But I think, this is more a driver issue than  
> Equalizer’s fault. The the additionally required  glReadPixels()/ 
> glDrawPixels() calls might also prevent optimal performance. What is  
> your experience here?

It depends on your base frame rate and data structure. As a rule of  
thumb, you don't get much speedup after ~20FPS. The pixel transfer is  
the major bottleneck on the Equalizer side. There are a couple of  
optimizations left to be implemented - namely PBO transfers and async  
readbacks. For PBO there is experimental code (see eqPixelBench), but  
so far I couldn't get it consistently to work faster than normal  
readbacks.

View frustum culling, fragment complexity and your data structures are  
an important factor as well.


> But again, I could not gain any performance increase. Is that even  
> possible with Equalizer (in the context of this config and hardware  
> setup) ? For AFR to work, the render-start of both pipes _must not_   
> be locked to each other, but instead, they should be offset by ca.  
> half a frame.
>
> So my question is: How to make DPlex work (optimally) on MGPU- 
> Machines?

You need one pipe (thread) for each channel (see 5-window.DPlex.eqc).

If your thread model is draw_sync (the default) you'll need one node  
(process) per GPU, otherwise your draws are serialized.

The statistics should show nice overlapping green draw operations  
between the two GPU's.


HTH,

Stefan.