The goal for the Intrinsically Resilient Overlay Network (IRON) and the Generalized Network Assisted Transport (GNAT) projects is improving the performance of networked applications that must exchange information over wide-area networks (WANs). The earlier IRON project focused exclusively on unicast applications; GNAT extended IRON to support multicast transports applications as well as applications that are latency sensitive.

Our core approach leverages fully distributed computation both within the network and at the edges that collectively works to continually maximize Cumulative Network Utility (CNU). CNU optimization leverages a robust, low-overhead, distributed optimization technique known as Backpressure, which is known to be throughput optimal. This work extends Backpressure by taking into account per-flow end-to-end throughput and latency requirements to create a system that is throughput optimal subject to latency constraints. The work is enhanced by the addition of a host of latency reduction techniques including latency sensitive hop-by-hop error control, congestion control, and a set of queuing delay reduction techniques to make end-to-end delays across the network as small as possible.

Backpressure itself is comprised of a pair of complementary techniques: a packet-forwarding algorithm that leverages queue differentials between neighboring overlay nodes to continually move traffic along the least congested paths; and a suite of admission control algorithms operating at the entry points to the overlay network that regulate access of individual application flows to the backpressure network.

The combined techniques work to continuously maximize network utility (and hence CNU) without requiring global knowledge of traffic pattern, packet priorities or network topology. Each GNAT node performs a local optimization that, in cooperation with its peers, yields a global maximization of CNU.