Enabling quick adaptation of network transport in evolving datacenter

Datacenters are evolving quickly with more diversified applications, more hardware accelerators, and new architectures (e.g., memory disaggregation). They pose significant challenges on existing network transport layers, because they not only generate heavier and more dynamic traffic load, but also require more stringent transport performance. Adapting the network transport to the quickly evolving datacenters is challenging: today operators often spend months on diagnosing transport anomalies and tuning/designing transport features, and this labor-intensive process will be worse with future evolution of datacenters. In this talk, I will show how to greatly simplify and speed up the process.

To make diagnosing existing transports (i.e., TCP) extremely intuitive, I design DETER, which can deterministically replay problematic TCP connections. To support replay, it just records very lightweight information during the runtime. Replaying TCP has unique challenges because switches are not replayable but they interact with TCP on a per-packet basis. The key idea is to record the interactions on an end-to-end, packet-stream basis rather than a per-switch, per-packet basis. DETER’s implementation based on Linux kernel is open-sourced, and has only 2~3% overhead of state-of-the-art diagnostic tools.

Going forward, an ideal transport protocol should automatically adapt to traffic changes. This is made possible by HPCC, a clean-slate protocol that utilizes new capabilities of new switches and NICs. Unlike existing protocols which use heuristics, HPCC precisely calculates the traffic rate that fully utilizes the bandwidth with near zero queuing. It achieves this with a new control law that precisely models the hardware nature, driven by the link load measurement from programmable switches. HPCC has been deployed in Alibaba Cloud and supported by most switch and NIC vendors.