next up previous contents
Next: Different topologies on local Up: GPU, a framework for Previous: Known problems with NAT   Contents

Download PDF version of this document Download GPU GPU project homepage


We sent the pi job (360'000'000,pi) to clusters of different sizes, composed by Intel Pentium III 930 MHz with 512 MB of RAM and running Windows XP SP1. The job was sent through the "simple frontend" in
/frontend/simple, so that the master computer was participating in the computation, too.

Jobs were computed in $ \mu=45.59$ seconds with a $ \sigma=2.87$. Variations were caused by random operating system load and GPU load. Tests with nodes outside the local network in the same city or abroad showed the same mean, but a higher sigma: the cluster was less homogeneous and there were faster and slower nodes.

To compute speedup, we simplify a little and plot the number of results received versus number of computers in a cluster.


Tiziano Mengotti 2004-03-27