Random Early Detection (RED) is the active queue management (AQM) algorithm recommended by IETF. Unfortunately, it is identified that RED is difficult to configure its parameters and the average queue length of RED is closely related to the load level. ARED (adaptive RED) is the adaptive version of RED. ARED dynamically adjust maximum packet marking probability according to the average queue length to make average queue length stable, but it still suffers from unstable instantaneous queue length and performance degradation under dynamic traffic conditions. In this paper, the cause of such problems of ARED is analyzed and a load adaptive active queue management scheme called LARED (load adaptive RED) is proposed. LARED features in adapting the load level of bottleneck link as well as quick response to queue length dynamics. Analysis and simulation results show that, compared with ARED and other AQM algorithms, LARED brings more stable queue dynamics; while keeping high link utilization and low queuing delay, it presents good responsiveness and robustness in various traffic conditions.