This paper develops a diffusion-approximation model for a stable GI/G/s queue: The queue-length process in the GI/G/s queue is approximated by a diffusion process on the nonnegative real line. Some heuristics on the state space and the infinitesimal parameters of the approximating diffusion process are introduced to obtain an approximation formula for the steady-state queue-length distribution. It is shown that the formula is consistent with the exact results for the M/M/s and M/G/• queues. The accuracy of the approximations for principal congestion measures are numerically examined for some particular cases.
