Conclusion

GPU shows with a prototype that distributed computing is feasible on a Peer-To-Peer network if we employ randomized algorithms. The plugin-frontend paradigm with GPU as a middle layer arises because we want all users not only to compute but also to send computational requests. A cooperative multithreading is used to display graphics only if needed: if the plugin window is not visible, we save CPU-time.

As an example, we implemented a plugin-frontend couple that solves a partial differential equation on an elliptic domain with a random walkers approach [1].

According to benchmarks, GPU scales up to about 64 computers in a particular 2D-grid configuration where each node has degree 4. In a random topology, the duplicates problem is still present and should be fixed in order to achieve the normal Gnutella scaling.

Some theoretical considerations are attempted: in particular, we discuss the coupon collector problem [27,28] and we give an estimation for the small world problem using fractal theory [24,25,26].

As final note, we remark that in most cases, it is not the lack of resources that stop us from using the CPU-power of so many idle computers, but the difficulty to implement parallel algorithms that produce meaningful results. Computations have a strong serial nature, indeed.