In a reservation network designed to carry bursty traffic an improved dynamic bandwidth allocation method accounts for reservation latency. In one advantageous embodiment wherein the system is memoryless, a user will request its queue size divided by the reservation latency in the system. Alternatively, a small memory based system can achieve further efficiencies by making available to a user a frame length amount of memory. The user can then request the greater as between its queue size divided by the reservation latency and the number of current arrivals at the user.