Lattice field theory on cluster computers: vector- vs. cache-centric programming

We evaluate the possibility of moving medium-sized calculation in lattice field theory from vector supercomputers to cluster computers, namely clusters built from Alpha processors and Myrinet interconnect, and find that a medium-sized system with a performance of 10 to 20 GFlop/s can be easily and cost-effectively built from current off-the-shelf components. The performance of the algorithms is analyzed with respect to memory bandwidth problems by experiment and using a cache simulator that uses C++ operator overloading. It seems that cluster systems, while hampered by poor memory bandwidth as compared to supercomputers, might offer opportunities for some algorithms that have good locality but are not vectorizable and thus will not perform well on vector systems