Traffic Management for AI-ML Data Centers

Traffic management is a reactive technique to handle traffic congestion as it occurs. One of the key types of traffic management for AI-ML data centers is priority-based flow control (PFC). Before configuring AI-ML traffic management features, configure class-of-service (CoS) features on your device. Read on to learn about the AI-ML traffic management features that Junos OS Evolved offers.

Table 1: General Traffic Management
Topic	Use Case
PFC Watchdog	Use the PFC watchdog to quickly detect and resolve PFC pause storms, maintain lossless traffic links, and improve link quality.
DSCP-based PFC for Layer 3 Untagged Traffic	You can configure DSCP-based PFC to support lossless behavior for untagged traffic across Layer 3 connections to Layer 2 subnetworks for protocols such as Remote Direct Memory Access (RDMA) over converged Ethernet version 2 (RoCEv2).
IPv6 Wildcard Mask Match Conditions	To dynamically apply firewall filters to similar IPv6 addresses, configure match conditions based on wildcard masks for your firewall filters.
Dropped Packet Notifications to Aid in System Performance Tuning	The dropped-packet notification feature enables you to see detailed information about what is causing particular packet drops. Having that information, in real time, allows you to tune up your system’s performance.
TAP Aggregation Enhancements	Test access point (TAP) aggregation provides N:M (any-to-any) packet replication, allowing you to capture different types of data in real time so that you quickly see what is happening in your network. Enhancements to the TAP aggregation feature provide timestamping and ACL filtering, as well as an updated hierarchy location for the TAP aggregation interfaces configuration.
Configuring BFD	The Bidirectional Forwarding Detection (BFD) protocol is a simple hello mechanism that detects failures in a network.
Drop Congestion Notification (DCN)	Drop Congestion Notification (DCN) is a congestion management technique based on packet trimming.
Stateless Flow Latency Monitoring	Stateless flow latency monitoring enables you to troubleshoot network flows that have unusually high travel time between their input and output interfaces on your switch.
DC Flow Analyzer	The DC Flow Analyzer (DFA) feature provides dynamic, hardware-assisted, per‑flow visibility on the switch. The DFA also helps track RDMA ROCEv2 flow destination queue pair information in Junos OS Evolved and other AI DC workloads.

Table 2: CoS Buffers
Topic	Use Case
Customize PFC X-ON Threshold and Per-Queue Alpha Values	Customize device responses to PFC pause storms to ensure that the device has enough time to resolve traffic congestion without disrupting your network.
Increase Shared Buffer Pool by Reducing Dedicated Buffer	Allow the device to allocate buffer space more efficiently among ports.

Table 3: ECN
Topic	Use Case
ECN Packets per Queue	Identify the packets that have experienced congestion to enable quick troubleshooting of network congestion points.
Latency-Based ECN	By aligning congestion signaling directly with acceptable queuing delay for each traffic class, you maintain integration with existing CoS scheduling and drop-profile behavior while expressing congestion tolerance in terms meaningful to latency-sensitive applications.