Non-minimal adaptive routing based on explicit congestion notifications

摘要

Low-diameter networks require non-minimal adaptive routing to deal with varying traffic characteristicsand avoid pathological performance. Such routing is based on local estimations ofnetwork congestion, based on link-level flow control credits. Dragonfly networks based on theextensions of commodity Ethernet networks using OpenFlow have been proposed for largeHPC deployments with low power consumption. However, this network technology does notimplement credit-based flow control. This work explores a range of routing solutions based onexploiting explicit congestion notification messages (in particular, 802.1Qau) to adapt the numberof packets using non-minimal paths. The design (denoted QCN-Switch) associates a probabilityvalue to each output port. This value is updated to reflect downstream congestion and used tostatistically divert traffic away from congested areas when the load is uneven, as in the case ofadversarial traffic.Afeedback comparison variant is designed to separate the cases of uniform trafficat saturation and adversarial traffic at low loads. Evaluation results show that QCN-Switch isa competitive design for both the uniform traffic and adversarial traffic. Furthermore, it is ableto react to changes in traffic conditions in 0.4 ms or less. A sensitivity analysis identifies the bestconfiguration and shows its performance trade-offs.