TCP at 100Gbit/s - Tuning, Limitations, Congestion Control

Link capacities increase at an enormous pace, with 100 Gbit/s
becoming standard in data centers, campus networks,and the
Internet. These ever increasing data rates are challenging since
end-system performance (esp. CPU performance) cannot keep up
with the growth rates. Still, the TCP protocol and today’s
hardware are capable of transferring 100 Gbit/s with a single
sender/receiver pair. However, extensive tuning is necessary down
to manual interrupt configuration and corresponding CPU core
pinning for the applications. A major issue is packet loss within
the receiving end-system that cannot be prevented by TCP’s flow
control. This, in turn, affects TCP’s default congestion control
that interprets the losses as congestion signal. In this paper
we show how to tune end-systems that are driven at their
performance limits, what data rates are feasible, where the
limitations are, and discuss the impact on and by TCP’s
congestion control.