A parallel algorithm for solving streamflow prediction problems using distributed models is proposed and analyzed. The expanding utilization of distributed hydrologic models for the realtime prediction of floods and flash floods and for hydroclimatic studies provides the impetus for the study. The efficiency of the proposed algorithm, as compared to ordinary sequential algorithms, as well as the detailed parametric influence on the performance of the algorithm are investigated. First, theoretical results for the speedup of the parallel algorithm are obtained for a simple stream network of a typical flash flood prone basin, and then they are extended to include a larger, well developed network of streams. In the latter case, the results are given as functions of geomorphological characteristics of the stream network and of the cost of generic procedures for soil water accounting and for channel routing. It is shown that when the number of parallel processors is not limiting, parallel processing is much more cost effective than sequential processing for stream networks of a high order. It is also shown that the algorithm performs better for higher bifurcation ratios and lower average-length ratios. Performance improves as the ratio of the computational cost of channel routing to that of soil water accounting is reduced. The algorithm is tested experimentally for a flash-flood prone basin in Ohio using an ENCORE parallel computer, yielding satisfactory results. Guidelines are provided for partitioning the drainage basins in a particular application, as well as for developing parallel software. (C) 1997 Elsevier Science B.V.