A flame length optimization scheme is proposed for multi-antenna downlink systems to guarantee diverse delay- bound violation probability constraints. Due to the difficulties of extracting the quality of service (QoS) metrics from the conventional physical-layer channel models, the link-layer models named effective bandwidth and effective capacity are applied to statistically characterize the source traffic patterns and the queuing service dynamics. With these link-layer models, the source traffic process and the channel service process are mapped to certain QoS parameters. The packet delay-bound violation probability constraints are converted into minimum data rate constraints and the optimization problem is thus formulated into simultaneous inequalities. With the assumption of ergodic block-fading channels, the optimal frame lengths of single-user and multiuser systems are calculated respectively by numerical iterative methods. Theoretical analyses and simulation results show that the given delay-bound violation probability constraints are well satisfied with the optimal frame length.