AN EXPERIMENTAL-STUDY OF PARALLEL BOLTZMANN MACHINE ON 2 DISTRIBUTED-MEMORY MULTIPROCESSORS

In our previous works1), 2), we have proposed an efficient mapping scheme of Boltzmann computations on a Distributed-Memory Multiprocessor (DMM) together with its expected speedup. In this scheme, the neurons in a Boltzmann Machine are partitioned into p disjoint sets and mapped onto each processor that simulates the mapped neurons in a timesharing fashion, where p is the number of processors in a DMM. This paper aims at applying the same parallelizing scheme to a Boltzmann Machine solving combinatorial optimization problems, and investigating the experimental speedups on widely used two DMM's: a network of Transputers and the Fujitsu AP1000 system. Experimental speedups show that there is a cost-effective number of processors in the parallel simulation with respect to the number of neurons in a Boltzmann Machine, such that though more processors are added to, the speedup ratio is not increased so much.