Finite-difference modeling of wave propagation in heterogeneous media is a useful technique in a number of disciplines, including earthquake and oil exploration seismology, laboratory ultrasonics, ocean acoustics, radar imaging, nondestructive evaluation, and others. However, the size of the models that can be treated by finite-difference methods in three spatial dimensions has limited their application to supercomputers. We describe a finite-difference domain-decomposition method for the three-dimensional acoustic wave equation which is well suited to distributed parallelization. We have implemented this algorithm using the PVM message-passing library, and show here benchmarks on two different distributed memory architectures, the IBM SP2 and a network of low-cost PCs running the Linux operating system. We present performance measurements of this algorithm on both the low-bandwidth PC network (10-Mbits/s Ethernet) and the high-bandwidth SP2 cluster (40-Mbits/s switch). These results demonstrate the feasibility of doing distributed finite-difference acoustic modeling on networks of workstations, but point to the substantial efficiencies that can be expected as higher bandwidth networks became available. (C) 1997 American Institute of Physics.