How UCloud’s DCN V4 Architecture Boosted Data Center Capacity and Performance

In October 2018 UCloud upgraded the basic network of its newly built data centers to the V4 architecture, moving all DCs to 25 Gbps/100 Gbps links, which quadrupled the per‑zone server port capacity to 320,000 and enabled lossless networking, horizontal scaling and rolling upgrades.

UCloud DCN V3 Architecture Design

UCloud’s public cloud provides services by zones (AZ), each composed of one or more data centers. The DCN V3, launched in 2016, featured a two‑level CLOS Spine‑Leaf design with 10 Gbps access and 40 Gbps inter‑connect, eliminating stacking, using BGP for spine routing, and supporting uniform deployment.

V3 addressed V2 limitations, offered horizontal scalability, and facilitated rapid overseas node deployment (Seoul, Tokyo, Washington, Frankfurt).

Challenges of V3

Rapid growth of GPU, real‑time big data, NVMe‑over‑Fabric and other workloads exposed performance, capacity and flexibility shortcomings:

Insufficient bandwidth and latency for high‑performance workloads.

Limited server port capacity (≈10‑20 k servers per DC) due to spine port count.

Uniform access model reduced flexibility for specialized network requirements.

DCN V4 Architecture Design and Optimization

To overcome these issues, the team redesigned the DCN, selected hardware, and standardized the solution, completing the V4 rollout in October 2018.

1. Hardware upgrade to 25 G/100 G platform

After extensive testing of over 300 product combinations, UCloud adopted a full 25 G/100 G hardware stack, enabling GPU zones to run on 10 G/25 G physical networks and doubling compute performance for AI training.

UCloud also deployed the new hardware in the Fujian GPU zone, achieving double the performance of the previous 10 G network.

2. 3‑level CLOS design

The new three‑level CLOS adds an aggregation layer (CS) between leaf (AS) and spine (DS). By increasing the number of spine devices, each CS consumes fewer spine ports, dramatically raising the total number of servers that can be connected.

The three‑level design can theoretically scale without bound; a single DC can now support up to 80 000 server ports, and a single zone up to 320 000 ports—four times the V3 capacity.

3. Introduction of PODs

A POD groups a set of aggregation switches, their downstream leaf switches and attached racks, providing a consistent network capability (e.g., uniform 1 × 100 G or 2 × 100 G links, 25 G server uplinks) and features such as ECMP, QoS, and direct Internet access.

Different business domains (public cloud, physical cloud, high‑performance big‑data, etc.) are mapped to specialized POD types (inner‑net POD, high‑performance POD, comprehensive POD, hybrid‑cloud POD) to balance flexibility and cost.

4. DC Group

DC Group connects geographically close DCs in a full‑mesh within the same AZ, reducing latency to <0.1 ms, increasing redundancy and bandwidth, and enabling seamless rolling upgrades without impacting running services.

Summary

UCloud’s DCN V4 architecture upgrades the physical network to 25 G/100 G, introduces lossless networking, expands capacity, and supports horizontal scaling and rolling upgrades, thereby reconciling new performance demands with legacy constraints and positioning the platform for several years of growth.