Credit-based Flow Control Based on RSVP for RDMA transmission in WAN
draft-hu-rtgwg-cbfc-rsvp-01

Abstract

This draft defines the Credit-based flow control mechanism for WAN based on the RSVP protocol. With the increasing demand for AI computing power, the computing power of a single AIDC can no longer meet the needs of large model training. This has given rise to cross-AIDC distributed model training, driving the demand for transmitting RDMA packets over WAN networks. AI training is extremely sensitive to network packet loss, and even a small amount of packet loss may lead to a significant decline in training efficiency. In addition, the elephant flow and extreme concurrent traffic also place higher demands on network performance. Credit- based flow control is a Backpressure-based traffic management technology, which has high reliability and stability in practical applications. It can provide high-throughput and zero-packet-loss transmission guarantees for RDMA traffic, effectively ensuring the efficiency of cross-data center AI training. This draft focuses on the scenario where RDMA packets are transmitted through SRv6 tunnels in the WAN and further expands the capabilities of the RSVP protocol in WAN. This draft introduces the Credit-based flow control mechanism into the RSVP protocol to achieve precise traffic control and provides processing analysis.

Authors

Jiayuan Hu
Zehua Hu
Tanxin Pi
Yongqing Zhu

(Note: The e-mail addresses provided for the authors of this Internet-Draft may no longer be valid.)