AI workloads exhibit traffic patterns that stress data center networks
unlike any other compute or storage workloads, posing unique challenges
for network performance and efficiency. In this talk, I will discuss
some of these challenges, and deep dive into one new mechanism that aids
traffic optimization and network efficiency in this context - a
practical and effective in-band signaling protocol called Congestion
Signaling (CSIG). We will explore CSIG’s use cases in congestion
control, traffic engineering and telemetry in software-defined networks
that serve AI traffic and beyond; and outline new avenues of research
that it enables.