eCore Architecture and Whitebox Evolution of Alibaba's Service-Oriented DCI Network

Author: Dennis (Alibaba, presenting for Royer); Manish (Cisco) Collection: 2025-OCP-Global-Summit-Networking

Summary

This talk describes Alibaba's eCore (ecore network), a next-generation service-oriented Data Center Interconnect (DCI) network that replaced their legacy global backbone (Aen) after a five-year development program. The design was motivated by the complexity, inflexibility, and operational burden of traditional backbone routers running MPLS, RSVP, and other legacy protocols. eCore's four design principles are: (1) Simplicity — single-chip single-stack, SRv6 + BGP only, ISIS restricted to national core nodes; (2) Composable architecture — pizza-box devices as Lego blocks, multi-plane and multi-space design for easy scaling; (3) Minimize blast radius — ISIS routing domain reduced by 2 orders of magnitude, fault isolation per plane; (4) Service-oriented — SRv6 VPN with 128-bit Service SID carrying both VPN identity and Traffic Engineering class in a single address.

The critical innovation is the SRv6 Service SID approach: instead of mapping DSCP values to RSVP-TE tunnels (limited to 64 DSCP values, operationally complex), eCore uses a 128-bit IPv6 address as the Service SID. This SID encodes both VPN identity (overlay) and TE class (underlay path selection) simultaneously, enabling per-prefix traffic engineering without any DSCP-to-tunnel binding. An endpoint (server, VM, container) simply places the Service SID in the packet; the eCore edge device reads the SID and routes to the appropriate SRv6 policy and TE tunnel automatically. eCore has been in production for approximately one year (two years since rollout started, 2025). The team of fewer than 10 engineers (hardware + software) developed the entire whitebox for eCore. A second-generation whitebox based on Cisco Silicon One P200 (51.2T single chip) is in development. Cisco's contribution covers the P200 digital twin (NSIM wrapped in a VSim) and programmable SBFD offload via P200's NPU host.