Aviz Networks details ONES 3.0 Fabric Manager orchestration for SONiC QoS

Aviz Networks’ ONES 3.0 Fabric Manager describes a SONiC QoS approach for AI switch fabrics, combining PFC and ECN with configurable scheduling. The update matters to enterprise architects and SOC and network teams managing GPU cluster traffic with low loss and congestion feedback.

Research Overview

The blog frames AI fabric performance as dependent on inter-server communication over a switch fabric, especially for GPU clusters running data-intensive workloads. It positions SONiC QoS features as the mechanism to manage lossless priority traffic and congestion behavior.

It links RoCE with Ethernet-based transfer efficiency and describes that preserving lossless, uncongested data streams requires QoS controls. It then ties those controls to orchestration via ONES 3.0.

Key Findings

The blog explains that packets with DSCP/DOT1P marking can be mapped to interface queues, and that enabling PFC on a queue provides lossless behavior for that priority. It describes how congestion detection in that queue triggers pause-frame signaling to stop the sender temporarily.

It also states that ECN uses buffer thresholds to send ECN-CNP packets to the sender once thresholds are exceeded, with the sender reducing transmission rate to avoid congestion. For egress, it describes choosing scheduling algorithms to prioritize traffic from higher-priority queues.

Technical Breakdown

For lossless transmission, the blog describes mapping incoming traffic markings (DSCP and DOT1P) to queues and enabling PFC for the selected queue. It adds that PFC pause frames are sent when congestion is detected, preventing packet drops for that priority class.

For congestion management beyond PFC, the blog describes ECN buffer thresholds and ECN-CNP packet signaling to senders. It then details egress scheduling options in SONiC, naming Deficit Weighted Round Robin (DWRR), Weighted Round Robin (WRR), and Strict priority scheduling (STRICT).

Operational Impact

The blog outlines how ONES 3.0 orchestrates SONiC QoS using YAML templates and automates creation and assignment of QoS profiles through Fabric Manager. It describes automated profile creation for DSCP-to-traffic-class mapping, traffic class to queue, and traffic class to priority groups, with profiles created using naming based on mapping values.

It further describes YAML-driven orchestration for PFC-enabled queues, including a PFC Watchdog configuration, and ECN configuration that creates a WRED_PROFILE attached to queues that are PFC enabled. It adds that scheduling policies are created when weights are specified for DWRR/WRR and bound to queues, and that ONES-FM can designate an ECN_CNP queue using STRICT scheduling so congestion notification packets have a queue path during heavy congestion.

Conclusion

The blog describes a SONiC QoS configuration model using PFC and ECN for lossless priority handling and congestion feedback, paired with scheduling controls and orchestration via ONES 3.0 Fabric Manager using YAML templates. It emphasizes that the same workflow supports Day-2 QoS changes through YAML template updates or the NetOps API, presenting a consistent approach for enterprise teams operating AI fabric networks. Blog Signals brief is a fact-based summary of the vendor blog.