Monte Carlo simulations of quantum systems on massively parallel supercomputers

A large class of quantum physics applications uses operator representations that are discrete integers by nature. This kind of application typically uses integer data representations and the resulting algorithms are dominated entirely by integer operations. An efficient algorithm for one such application has been implemented on the Intel Touchstone Delta and iPSC/860. The algorithm uses a multispin coding technique which allows significant data compactification and efficient vectorization of Monte Carlo updates. The algorithm regularly switches between two data decompositions, corresponding naturally to different Monte Carlo updating processes and observable measurements such that only nearest-neighbor communications are needed within a given decomposition. On 128 nodes of Intel Delta, this algorithm updates 183 million spins per second (compared to 21 million on CM-2 and 6.2 million on a Cray Y-MP). A systematic performance analysis shows a better than 90% efficiency in the parallel implementation.