Aviz Networks details Subscriber-Aware Load Balancing on NVIDIA BlueField-3

Aviz and NVIDIA describe Subscriber-Aware Load Balancing on the NVIDIA BlueField-3 DPU, designed to correlate control- and user-plane traffic per subscriber at line rate. The approach targets accurate subscriber-level analysis for deep network observability while reducing CPU overhead.

Research Overview

The post frames a challenge for mobile operators: control-plane packets such as GTP-C, PFCP, and N11 and user-plane packets such as GTP-U can traverse separately through the core. Without subscriber-level correlation, operators risk losing context and reducing efficiency in performance tools.

To address this, Aviz Service Node includes Subscriber-Aware Load Balancing (SALB) running on NVIDIA BlueField-3. The post states that SALB provides traffic steering and real-time correlation while maintaining throughput at line rate.

Key Findings

Aviz and NVIDIA report testing in which BlueField-3 delivered line-rate 100 Gbps throughput while using one Arm core and the P4 datapath, compared with roughly 20 x86 CPU cores needed for the same performance. The test uses GTP user and control plane traffic.

The post also presents power comparisons for scenarios at 100 Gbps and 400 Gbps, reporting BlueField-3 power of about 110W versus COTS x86 power of about 160W for the 100 Gbps case. For the 400 Gbps scenario, it reports about 110W for BlueField-3 versus about 560W for COTS x86.

Technical Breakdown

SALB is described as correlating control- and user-plane packets for each subscriber so both traffic types are processed together. The goal is subscriber-coherent steering that supports analytics while avoiding additional CPU strain.

On BlueField-3, the post states that control-plane packets are processed on the DPU Arm cores, while subscriber context is extracted for user-plane handling. User-plane traffic steering is implemented using the DPU’s P4 pipeline for line-rate processing.

Architecture and datapath behavior

The post describes the DPU mode as configured for running the Aviz Service Node, with control-plane processing on the Arm cores and user-plane flows offloaded to the P4 pipeline. It also states that ASN uses the NVIDIA P4 DPL SDK to install rules proactively for fast-path processing.

For operators, the post links this setup to balanced probe capacity, preservation of traffic fidelity, and visibility maintained per subscriber while steering flows at line rate.

Operational Impact

The post connects SALB to three operational outcomes: subscriber-coherent analytics, improved tool efficiency through probe load balancing, and performance per watt via DPU offload. It describes the feature as sending both control- and user-plane packets for a subscriber to the same tool.

For resource planning, the post reports replacing tens of x86 cores with DPU-based handling and cites power savings aligned with those core reductions. It also notes joint validation showing SALB can run on a single BlueField-3 DPU using its P4 programmable datapath and Arm cores.

Overall, the blog describes a Subscriber-Aware Load Balancing approach that correlates control- and user-plane packets per subscriber on NVIDIA BlueField-3, with reported line-rate throughput and lower CPU and power requirements versus x86-based processing. Blog Signals brief is a fact-based summary of the vendor blog.