An I/O management method includes determining local target I/O rates for a plurality of storage controllers across which an I/O stream is distributed. The local target I/O rates may reflect a target latency corresponding to a QoS attribute associated with the I/O stream. A cumulative target I/O rate may be determined based on the local target I/O rates. A token bucket rate may be calculated for each applicable storage controller based on the local target I/O rates. The storage controllers may implement token bucket algorithms in accordance with the token bucket rates to enforce the QoS for the I/O stream. The local target I/O rates may be determined periodically based on measured data indicating actual latency and traffic. A queue depth is determined based on the measured values of latency and traffic. The queue depth is then used with the desired value of latency to obtain the target I/O rates.