draft-ruan-spring-priority-flow-control-sid
SRv6 behavior extention for Flow Control in WAN
In the wake of the continuous emergence of novel AI technologies,
including collaborative training, distributed inference and the like,
the scenario in which different data centers communicate via the RDMA
(Remote Direct Memory Access) protocol has emerged as a new
requirement. similar to data center networks, wide area networks
(WANs) should also be equipped with certain flow control
capabilities, nodes in the wide area network (WAN) need to quickly
and accurately notify upstream traffic nodes of their congestion
status. Upon receiving the congestion notification, upstream nodes
can take responsive actions to mitigate network congestion. By
extending SRv6[RFC 8986] technology, specific services can be carried
over SRv6 policies, which enables the precise delivery and response
of congestion signals.
For traffic carried over SRv6 policies, the service traffic path is
explicitly orchestrated via the Segment List in the SRH header. By
parsing the sequence of the Segment List, downstream nodes or
controllers can accurately identify the upstream nodes and interfaces
corresponding to the current traffic, thus enabling the rapid
notification of congestion information to the upstream. Since each
SID in the Segment List is a 128-bit IPv6 address that represents a
specific behaviour (e.g., End, End.X), extending existing behaviours
to mark their capability to receive and process congestion signals
allows flexible control over the delivery mode of congestion signals
in the WAN.
The specific packet formats and contents of congestion signals vary
in various ways and are continuously being updated. This document
takes the most commonly used PFC backpressure notification mechanism
as an example to elaborate on the overall workflow of WAN congestion
information notification and the extensions to the SRv6 protocol.
PFC technology is widely deployed in RoCEv2 networks within data
centers to report congestion signals. By leveraging SRv6 and NRP
technologies, it is possible to effectively control the propagation
range and path of PFC backpressure signals in wide area networks,
thereby avoiding issues such as deadlocks and the excessive
propagation of congestion scope.