Opportunistic Packet Forwarding for Proactive Transport in Datacenters

Proactive transport protocols in datacenters are designed to avoid congestion by regulating flow sending rates via credit allocation. However, when a new flow starts, it takes one RTT before credits can be assigned to the new flow. To avoid stalling flows, modern proactive protocols such as NDP and Homa allow a new flow to blast a burst of unscheduled packets at line rate during the pre-credit phase. However, sending too many unscheduled packets could cause temporary traffic spikes that lead to queue build-ups, packet losses, and retransmissions, which are particularly detrimental to short flows. In this paper, we present the design and evaluation of Opportunistic Packet Forwarding (OPF), a data-plane building block for proactive transports designed to minimize pre-credit packet losses with negligible overhead on network switches. The key idea in OPF is to allow pre-credit packets to opportunistically take detours to avoid congested links on the shortest paths, effectively trading off packet losses for slightly increased delay. We have implemented OPF using P4 switches and integrated our implementation with both Homa and NDP. Our results on a range of traffic loads show significant improvement in the 99-th percentile of flow completion time for short flows, namely up to 54% reduction in NDP and 50% in Homa.