Inside Alibaba Cloud’s HAIL Network: Architecture, Innovations, and Future Trends

What is HAIL?

HAIL is the codename for Alibaba Cloud’s data‑center network architecture, standing for Highly Availability, Intelligence, and Low‑latency , reflecting the core goals of high availability, intelligent control, and minimal delay in the cloud network.

Evolution of Data Center Networks

Early data‑center networks were built on enterprise‑grade equipment (VPC, stacking, OSPF) and served simple intra‑datacenter communication. As cloud services grew to serve global users, the need for massive concurrent processing, fast storage, and high‑speed interconnects pushed networks toward specialized, large‑scale designs, exposing performance, stability, and operational challenges.

Alibaba’s Self‑Developed Network (HAIL DC5.2)

Since 2013 Alibaba moved from standardized commercial gear to a fully self‑designed stack, culminating in the 2017 HAIL architecture.

HAIL DC5.2 introduced a single‑chip, box‑type switch that enables multi‑plane scale‑out, simplifying hardware/software complexity and focusing development on network stability.

Decades of operational experience were codified into the NET system platform, automating large‑scale deployment, monitoring, and fault response.

Fully self‑designed hardware and software are tightly integrated with backend monitoring, delivering high‑precision, real‑time performance metrics and automated remediation.

Current Architecture and Technologies

From 2019 onward, new data centers adopt the AliNOS‑based self‑developed switches covering campus core, cluster core, POD core, TOR access, and P4‑programmable gateway devices (SNA).

Multi‑plane interconnect supports flexible scaling from thousands to hundreds of thousands of servers using a three‑tier CLOS design.

Scale‑out redundancy makes single‑device failures virtually invisible to the overall datacenter.

High‑radix single‑chip switches reduce hop count while maintaining massive scale, compressing internal forwarding latency.

Two device families cover all interconnect scenarios, cutting supply and operational marginal costs.

Eliminating traditional stacking with dual uplinks removes stability risks and enables seamless software upgrades, crucial for rapid feature iteration.

Future Technical Trends

Physical chip throughput will keep rising as network‑chip Moore’s law continues, driven by heterogeneous compute demands and cost reduction pressures.

Ultra‑low‑latency forwarding becomes essential for HPC, AI, storage, and database workloads, requiring stack‑wide innovations from protocols to flow control.

Pooling – both large‑scale compute/storage pooling and fine‑grained resource pooling – will push networks to deliver predictable, low‑latency IO while scaling linearly.

Predictable network concepts aim to give applications reliable expectations of network behavior, simplifying application architecture much like high‑speed rail offers consistent travel times.

Predictable Network Vision

Alibaba Cloud envisions a fully self‑controlled stack—custom NICs, switches, optical links—and a collaborative design approach that aligns network and application layers, end‑side and exchange‑side designs, and architectural upgrades. Recent conference announcements highlighted new products built on this philosophy, integrating autonomous hardware, high‑performance protocols, programmable acceleration, and intelligent operations into a cohesive, predictable data‑center network system.