This paper evaluates the effectiveness of Random Early Detection (RED) in sharing bandwidth fairly among different traffic types, including non-adaptive, fragile, and robust flows. RED can lead to unfair bandwidth sharing when multiple traffic types share a link, as it applies the same loss rate to all flows regardless of their bandwidths. To address this, the authors propose Fair Random Early Drop (FRED), a modified version of RED that uses per-active-flow accounting to impose a loss rate based on each flow's buffer use. FRED provides better protection for adaptive flows and isolates non-adaptive traffic more effectively. It also works with FIFO gateways and uses memory proportional to the total buffer usage, maintaining state only for flows with buffered packets. FRED improves fairness by keeping state for only those flows that have packets buffered in the gateway. This approach reduces the cost of per-active-flow accounting, which is proportional to buffer size and independent of the total number of flows. FRED can be extended to support gateways with large numbers of buffers, which helps in managing high numbers of active flows. A "two-packet-buffer" mechanism is introduced to allow the use of large numbers of buffers without incurring additional queueing delays. The paper analyzes the behavior of RED with different traffic types, showing that proportional packet dropping does not always lead to fair bandwidth sharing. RED can be biased against fragile connections and cause unfairness in bandwidth allocation. FRED addresses these issues by selectively dropping packets based on buffer usage, providing better protection for adaptive flows and more effective isolation of non-adaptive traffic. Simulations show that FRED is fairer than RED in handling connections with different round trip times and window sizes, and it protects adaptive flows from non-adaptive flows by enforcing dynamic per-flow queueing limits. FRED is also compatible with existing FIFO queuing architectures and can be configured to operate in byte mode. The paper concludes that selective discard mechanisms are needed to protect flows using less than their fair share and prevent aggressive flows from monopolizing buffer space and bandwidth.This paper evaluates the effectiveness of Random Early Detection (RED) in sharing bandwidth fairly among different traffic types, including non-adaptive, fragile, and robust flows. RED can lead to unfair bandwidth sharing when multiple traffic types share a link, as it applies the same loss rate to all flows regardless of their bandwidths. To address this, the authors propose Fair Random Early Drop (FRED), a modified version of RED that uses per-active-flow accounting to impose a loss rate based on each flow's buffer use. FRED provides better protection for adaptive flows and isolates non-adaptive traffic more effectively. It also works with FIFO gateways and uses memory proportional to the total buffer usage, maintaining state only for flows with buffered packets. FRED improves fairness by keeping state for only those flows that have packets buffered in the gateway. This approach reduces the cost of per-active-flow accounting, which is proportional to buffer size and independent of the total number of flows. FRED can be extended to support gateways with large numbers of buffers, which helps in managing high numbers of active flows. A "two-packet-buffer" mechanism is introduced to allow the use of large numbers of buffers without incurring additional queueing delays. The paper analyzes the behavior of RED with different traffic types, showing that proportional packet dropping does not always lead to fair bandwidth sharing. RED can be biased against fragile connections and cause unfairness in bandwidth allocation. FRED addresses these issues by selectively dropping packets based on buffer usage, providing better protection for adaptive flows and more effective isolation of non-adaptive traffic. Simulations show that FRED is fairer than RED in handling connections with different round trip times and window sizes, and it protects adaptive flows from non-adaptive flows by enforcing dynamic per-flow queueing limits. FRED is also compatible with existing FIFO queuing architectures and can be configured to operate in byte mode. The paper concludes that selective discard mechanisms are needed to protect flows using less than their fair share and prevent aggressive flows from monopolizing buffer space and bandwidth.