Routers and networks with near-zero buffers

Routers in the Internet today employ buffers to store packets during periods of congestion. With Internet traffic expected to grow phenomenally over the next five to ten years, and as core link speeds approach terabits-per-second, it is becoming increasingly difficult (due to energy/power consumption and other technological hurdles) to equip high-speed routers with large electronic buffers comprising of SRAM (static) and DRAM (dynamic) chips. It appears that the most promising way forward to accomplish the capacity scaling and to reduce power consumption in the core is to increase the role of optics in packet switching. However, building an on-chip integrated all-optical buffer to store even a few packets is by no means a simple task. These considerations have engendered a flurry of debates regarding the size of buffers required at core Internet routers. Recent arguments supported by theory and experimentation show that under certain conditions, core router buffer sizes of a few tens of KiloBytes suffice for realising acceptable end-to-end TCP throughputs.In this thesis, we first trace the evolution in thinking over recent years on how much buffering is required at Internet routers, focusing specifically on the push towards smaller buffers, from GigaByte down to KiloByte sizes, making them amenable for all-optical realisation in the future.Second, as the popularity of real-time applications grows in the Internet, we investigate the interaction that happens when closed-loop TCP traffic multiplexes with open-loop (i.e., real-time) UDP traffic at a router with very small (optical) buffers, and make a significant discovery. Contrary to conventional wisdom, we show that in the regime of very small buffers (up to 50 KiloBytes), losses for real-time traffic do not fall monotonically with increasing buffer size. Instead, they can exhibit an anomalous performance wherein larger buffers induce higher losses.Third, we explore if the anomalous behaviour can be alleviated by dedicating (i.e., pre-allocating) buffers to UDP traffic. Our results indicate that while dedicating buffers to real-time traffic indeed mitigates the anomaly, determining the amount of buffers that should be pre-allocated to UDP and TCP traffic is non-trivial.Fourth, we envision a near-zero buffer core optical network, and propose a novel edge-to-edge based packet-level forward error correction (FEC) framework as a means for recovering the lost packets in the near-zero buffer core. We consider a realistic multi-hop network, develop a global optimisation framework, and design a practical heuristic algorithm to ensure network-wide fairness (in terms of average per-flow goodput and flow completion times) between single- and multi-hop flows. We show that our FEC framework, when tuned properly, can be very effective in overcoming high core losses, thus paving the way for an all-optical bufferless core network in the future.Finally, we profile how TCP reacts to the reduced buffer availability in the network by studying the burstiness of the traffic arrival process to a bottleneck link. We show that while large buffers can create significantly bursty traffic, small buffers render the TCP aggregate to behave more like a Poisson process. This observation can have a huge impact on the design of a future small buffer core network.The work in this thesis highlights the various factors that need to be considered when provisioning routers and networks with near-zero buffers, presents solutions that address these factors, and represents a step towards the realisation of a wide-area bufferless core optical network