Dual Channel Adaptive Congestion Control for Datacenters

An ideal datacenter network congestion control mechanism should provide zero packet loss, full link capacity utilization, zero buffer occupancy as well as fast fairness convergence. Most of the existing congestion control protocols employ the responsive control approach (control after congestion occurs), which are hard to simultaneously satisfy these conflicting goals. Addressing this issue, this paper presents a novel end-to-end datacenter congestion control approach, named SCP (Stand-in Control Protocol). Different with the existing solutions, SCP realizes the preventive congestion control – control before congestion occurs. SCP configures and logically partitions a physical link into control and data channels. Before sending real data packets in the data channel, SCP sends stand-in packets in the control channel to detect the path condition. Based on the preview results of the control channel detections, SCP ensures that the data channel is never congested yet approaches maximum throughput; thus zero packet loss, full utilization and zero buffer occupancy are simultaneously satisfied in the data channel. Moreover, SCP employs an adaptive online learning rate adjustment algorithm to achieve fast convergence. The adaptive learning algorithm can be automatically adjusted to any datacenter environment without any fundamental assumptions. Extensive experiments shows that SCP outperforms some existing lossless congestion control protocols and SCP's novel design makes itself can be applied not only in the datacenters but also in the WAN (Wide Area Network) scenarios.