DNN training in geo-distributed data centers using VXLAN and EVPN

Abstract

The exponential growth of data and increasing demands for privacy and compliance drove the shift toward geo-distributed training of Deep Neural Networks (DNNs). Traditional centralized training in a single data center faced limitations such as WAN bandwidth constraints, high latency, and cross-border data restrictions. Conventional VLAN-based approaches also lacked scalability for multi-tenant, distributed environments due to limited ID space and poor isolation. To overcome these challenges, this thesis designed and evaluated a scalable, resilient, multi-tenant virtual network architecture for geo-distributed DNN training. The system used VXLAN as a Layer 2 overlay and EVPN over MP-BGP as the control plane to extend network reachability across data centers while enabling isolated virtual networks. A realistic emulated environment was built using Containerlab and FRR, with a spine-leaf topology modeling inter-data center communication. The network was enhanced with Equal Cost Multi Path (ECMP) routing for load balancing and Bidirectional Forwarding Detection (BFD) for fast failure recovery. Prometheus, combined with SNMP and ping exporters, enabled continuous monitoring of link health, device uptime, and overall network connectivity.