Our network connection from the Visualization Lab to the Paragon is through a shared ethernet link. Sending commands from the workstation and bringing images back from the Paragon is a slow process. The average (round) message latency is around 16 ms (it varies from 9 ms to 160 ms). (This was measured when the network was lightly loaded.) This latency could be acceptable for real-time manipulation if it took almost no time to render and transfer the images back. But the reality is that rendering and transferring the images take much longer than the minimum necessary for real-time manipulation. Push-button interactivity might be achievable depending on other factors.
Getting the images back over the ethernet is a problem. First, the
link between our laboratory and the machine is shared and we see
longer delays during the day than at night. Also, we should avoid
using all of the available bandwidth because of other users. The
throughput varies from 28KB/sec (for 100 byte messages) to 108KB/sec
(the maximum ethernet payload of about 1500 byte message). Without
compression we can only transfer a couple of 
gray scale images
at a time. Fortunaly, compression techniques are available, but to use
these we must perform compression on the fly, which is very
computationally expensive.
Figure 10: Latency for compositing images of different
sizes on the Intel Paragon.
Figure 11: Throughput for compositing images of
different sizes on the Intel Paragon.