In future wavelength-division multiplexed networks, wavelength resources are efficiently utilized by statistical multiplexing and dynamic lightpath provisioning. In order to provide desired blocking probability and the ease of quality of service provisioning, this paper concentrates on the end-to-end connection acknowledged optical burst switching (OBS) protocols, and proposes a prebooking mechanism to improve the lightpath bandwidth efficiency, which consequently, reduces the blocking probability and better supports latency-sensitive traffic. The prebooking mechanism excludes lightpath acknowledgement delay from the lightpath holding time by "prebooking" each aggregated burst/flow based on the estimated characteristics of arriving traffic. More specifically, a burst length prediction and reservation strategy is proposed and analyzed for the widely observed self-similar incoming traffic. The simulation results show that the proposed excessive reservation strategy is able to achieve low bit loss caused by the prediction error. In addition, the proposed mechanism is also investigated with constant bit rate traffic in a large mesh backbone, such as NSFNET. The results demonstrate that in large networks, the proposed prebooking mechanism is able to bring significant benefits in terms of blocking probability and lightpath bandwidth efficiency. For example, with a maximum acceptable end-to-end delay of 90 ms and a 10 -4 bit loss tolerance, our prebooking mechanism yields approximately double the supported traffic load relative to the familiar wavelength-routed OBS architecture