Priority Policy in Multi-Queue Data Center Networks via per-Port ECN Marking

摘要

Explicit Congestion Notification (ECN) as a main part of the Internet congestion control was originally designed with a single queue in mind. But, nowadays data centers industry trends to have multiple queues per-port in its switches. This type of mismatch can affect the scheduling schemes and hence leads to remarkable performance degradation in multi-queue switches. For instance, when some of the queue buffers of the switch exceed the considered thresholds, all the packets belonging to the same shared port could get ECN mark. To cope with this challenge, we propose Priority-ECN, a first systematic solution to this problem. The first intuition of Priority-ECN lies in the fact that if any packet gets marked, it should be prioritized to pass a route quicker than the others to notify the condition of the network in an early stage. This presents a full view of instantaneous demand from all senders. The second intuition is that if any packet gets marked in the queue buffer, Priority-ECN will not consider them in the marking decision again in the output port buffer. The third intuition is to reach the goal of high throughput without sacrificing the latency. So, we use an approach similar to Cut Payload (CP), which drops the payloads of packets, rather than the metadata, when a queue reaches the threshold. We do so through dual couple buffer in output port and catch the advantage of CP to implement an efficient scheme to obtains near optimal flow completion times across different flow sizes with very low latency.