Aviz Networks details ONES 2.0 for SONiC AI fabrics RoCE monitoring

Aviz Networks’ ONES 2.0 adds RoCE-focused telemetry and GUI visibility for SONiC-based AI fabric networks, including PFC counter support and congestion management tied to port and queue utilization. For enterprise operators, the update centers on monitoring lossless Ethernet behavior and queue-level performance metrics that affect low-latency AI workloads.

Research Overview

The article frames generative AI network requirements as a set of operational needs: predictive analytics, QoS enhancements, resource optimization, anomaly detection, simulation of realistic environments, and autonomous operations. It also ties these needs to RoCE (RDMA over Converged Ethernet) as an approach for improving training-relevant communication characteristics on Ethernet networks.

It further describes the role of an AI fabric as the basis for improving model training speed, optimizing data movement, and leveraging compatibility with Ethernet networks. The article positions traffic monitoring in RoCE environments as a factor in maintaining operating continuity.

Key Findings

The article states that ONES 2.0 collects RoCE monitoring metrics to support flow control visibility, with a focus on understanding traffic prioritization and congestion behavior. It links monitoring to identifying congestion points and bottlenecks in advance of performance degradation for AI workloads.

The document also describes proactive congestion management as a way to reduce performance degradation when AI workloads exchange large datasets and real-time node-to-node communication. It states that PFC supports pausing non-critical traffic during congestion to limit packet loss and protect RDMA communication.

Technical Breakdown

According to the article, ONES 2.0 includes Priority Flow Control counters for RoCE support, plus congestion management based on port and per-port queue utilization details. The described metric set includes PFC counters, Rx/Tx watermark counters, and QoS drop counters, with queue drop counters used to track dropped packets inside the queuing system.

For operational monitoring, the article describes how metrics support traffic prioritization, congestion management, QoS adherence, identification of bottlenecks via pause-frame behavior, and real-time monitoring for responsiveness to network condition changes. It also describes performance optimization, capacity planning contributions, and queue-level insights used to adjust configurations such as traffic prioritization and resource allocation.

Product Update

The article presents ONES 2.0 as a release for SONiC-based AI fabrics that integrates monitoring and management features through an “ONES ecosystem” that includes orchestration and visibility. It says the platform supports third-party APIs including REST and Prometheus and is intended to collect metrics for SONiC-Fabrics across multiple vendor platforms.

For visualization, it describes GUI views showing RoCE traffic topology between RoCE communicating devices, interface-level indicators for PFC-enabled and RoCE-capable interfaces, and device-specific provisions for RoCE support such as L3 lossless traffic handling on selected queues. It also describes a view for RoCE traffic segregation alongside regular traffic and a display for pause-frame statistics.

Blog Signals brief is a fact-based summary of Aviz Networks’ vendor post describing ONES 2.0 for SONiC-based AI fabrics, including RoCE telemetry (PFC counters, Rx/Tx watermarks, QoS drop and queue drop counters), proactive congestion management, and GUI views plus REST and Prometheus integrations.