This paper introduces a new multiscale framework for estimating the tail probability of a queue fed by an arbitrary traffic process. Using traffic statistics at a small number of time scales, our analysis extends the theoretical concept of the critical time scale and provides practical approximations for the tail queue probability. These approximations are non-asymptotic; that is they apply to any finite queue threshold. While our approach applies to any traffic process, it is particularly apt for long-range-dependent (LRD) traffic. For LRD fractional Brownian motion, we prove that a sparse exponential spacing of time scales yields optimal performance. Simulations with LRD traffic models and real Internet traces demonstrate the accuracy of the approach. Finally, simulations reveal that the marginals of traffic at multiple time scales have a strong influence on queuing that is not captured well by its global second-order correlation in non-Gaussian scenarios.