First test was run using GPU 0.80x at the end of October in the room HG E27. Results clearly show that GPUs in this version do not scale linearly: the number of received results is growing exponentially in one of the three series; in that series, we found one crashed GPU. In the other series, we occasionally had a number of results greater that the number of computers involved, this was also incorrect.
A second test at the beginning of January failed because of a router crash in IFW C31. As support people at ETH said, GPU was not involved in the crash, fortunately.
We run a last test with version 0.847. To reduce a little the duplicates problem, we chose a square grid where each node is connected to 4 neighbors. We subsequently added nodes, 21 were on a local network at ETH, Zürich, one in a student's house in the same city and one was in Russia from a user who occasionally connected to the network.
In this particular 2D-grid configuration, GPU scaled quite well; sometimes we had more results, sometimes less though there was no exponential growth of packets. GPU should scale up to a 8x8 grid with 64 computers. Larger grids are not possible because of the TTL constraint we saw in the line topology.
Figure 20: First big cluster on a random graph (January, version 0.80x).
Figure 21: Second big cluster test on a square grid (March, version 0.847)