A selective packet discard mechanism is disclosed for selectively discarding packets at buffers in packet switches in the event of traffic congestion. The mechanism makes use of ordering information contained in data packets, such as forward sequence numbers or timestamps, to estimate the packet loss that traffic flows may have already incurred at upstream packet-switching nodes. This estimated upstream packet loss information is used to make improved packet discard decisions. The mechanism may be applied independently to a plurality of buffers at a plurality of packet switches to form a distributed data traffic management system.