|
This video is part of the appearance, “Juniper Networks Presents at AI Infrastructure Field Day 2“. It was recorded as part of AI Infrastructure Field Day 2 at 08:00 - 11:30 on April 23, 2025.
Watch on YouTube
Watch on Vimeo
Juniper Networks presented its latest Apstra functionality for AI data center network operations at AI Infrastructure Field Day. It focused on providing operators with the context and tools to manage complex AI networks efficiently. Jeremy Wallace, a Data Center/IP Fabric Architect, emphasized the importance of context in understanding the network’s expected behavior to identify and resolve issues quickly. Juniper is leveraging existing Apstra capabilities, augmented with new features such as compute agents deployable on NVIDIA servers, and enhanced probes and dashboards, to monitor AI networks. This presentation aims to equip operators to maintain optimal performance and minimize downtime in critical infrastructure environments.
The presentation highlighted the evolution of network management for AI data centers, transitioning from traditional methods to a more proactive and data-driven approach. The core of Juniper’s solution involves leveraging telemetry, including data collected from GPU NICs and switches, to provide real-time insights into network performance. This enables operators to monitor key metrics, such as GPU network utilization and traffic patterns, and respond to potential issues swiftly. The Honeycomb view, traffic dashboards, and integration with congestion control mechanisms (ECN and PFC) demonstrate how to provide visibility into the network’s behavior. The goal is to provide context and the tools to diagnose and resolve problems faster.
Finally, Wallace demonstrated a live demo of the platform, showcasing features like real-time traffic analysis, heatmaps of GPU utilization, and auto-tuning load balancing. The auto-tuning functionality dynamically adjusts parameters like inactivity intervals to optimize performance and eliminate out-of-sequence packets, increasing the likelihood of successful job completion. These power packs are essentially Python scripts and are evolving, with Juniper actively working on creating more of these power packs. Juniper is also working on deeper integration with other vendors for their customers’ environments and solutions.
Personnel: Jeremy Wallace