Technology
for Optical Pixel Streaming
As stated in the OptIPuter principle, the bandwidth of modern networks exceeds
the bandwidth of the internal bus of a traditional computer. If we ignore latency,
it is possible to geographically distribute the computer so that disks are in
Tokyo, graphics card in Amsterdam and display in New York.
SARA has been studying the separation of graphics card and display in this concept.
It must be noted that when we mention an OptIPuter graphics card or an OptIPuter
display, that we are refering to a large-scale infrastructure. Thus an OptIPuter
display is actually a cluster of displays, or a tiled panel display. And an OptIPuter
graphics card is actually an aggregate of high performance workstations, each
equipped with fast 3D hardware and dual processors.
SARA's approach to driving tiled displays with remote render clusters is a lightweight
approach. Our key insight into this problem is based on the short lifespan of
a picture element in a video stream. At 50 frames per second, a picture element,
or pixel, is only visible for 20 ms. If such a pixel is lost during communication,
chances are it will remain unnoticed, because apart from spatial coherence, video
streams also exhibit temporal coherence. Our approach therefore uses unreliable
network transport (UDP) as opposed to a reliable (TCP) transport. This will give
a valuable benefit, as high throughput is easily realized when acknowledgements
from remote sites with large latencies are no longer required.
The viability of our approach was successfully demonstrated during SuperComputing
2004 exhibition where high-resolution images were streamed live from Amsterdam
to Pittsburgh over a 10Gb/s line, whereby the application consumed 6.4 Gb/s bandwidth.
The application was controlled with a game pad at display side, and the low bandwidth
control signals where sent reliably over TCP to Amsterdam. The high-bandwidth
return signal represented the flood of high-resolution imagery (6400x4800 pixels
per frame).
For iGrid 2005, we intend to scale up. A sub-project of Jason Leigh's Global Lambda
Visualization Facility we will be testing the approach, not with the previous
16 panel setup, but with EVL's behemoth of 11x5 panels. The display, at CalIT^2
with UCSD in San Diego, will be bombarded with this unreliable stream, of what
we hope, will be 20 Gigabit/s of images, live from Amsterdam. We will be streaming
both 2D and 3D content to the 100Mpixel panel of EVL.
Additionally, we intend to do some comparative studies with EVL's approach to
scalable graphics, called SAGE. SAGE is a much more flexible framework, with more
possibilities, such as multiple video streams all scalable on a single tiled display.
We intend to use SAGE to stream live content from Amsterdam to the tiled display
as well.