ECN Marking With Micro-Burst Traffic: Problem, Analysis, and Improvement

In data centers, batching schemes in end hosts can introduce micro-burst traffic into the network. The packet dropping caused by micro-bursts usually leads to severe performance degradations. Therefore, much attention has been paid to avoiding buffer overflow caused by micro-burst traffic. In particular, ECN is widely used in data centers to keep persistent queue occupancy low, so that enough buffer space can be available as headroom to absorb micro-burst traffic. However, we find that current instantaneous-queue-length-based ECN marking scheme may cause problems in another direction - buffer underflow. Specifically, current ECN marking scheme in data centers is easy to trigger spurious congestion signals, which may result in the overreaction of senders and queue length oscillations in switches. Since ECN threshold is low, the buffer may underflow and link capacity is not fully used. In this paper, we reveal this problem by experiments. Besides, we theoretically deduce the amplitude of queue length oscillations. The analysis results indicate that the overreaction of senders is caused by ECN mismarking. Therefore, we propose combined enqueue and dequeue marking (CEDM), which can mark packets more accurately. Through test bed experiments and extensive ns-2 simulations, we show that CEDM can significantly reduce throughput loss and improve the flow completion time.