Why Hyper‑Converged Data Center Networks Are the Future of AI‑Driven Infrastructure

Why Hyper‑Converged Data‑Center Networks Are Needed

Modern data‑center workloads, especially AI, require the network to handle three traffic classes—general compute, storage, and high‑performance‑computing (HPC)—on a single fabric. Traditional deployments use three separate networks (InfiniBand for HPC, Fibre Channel for storage, Ethernet for general compute), which leads to higher latency, complexity, and cost.

AI‑Era Drivers

SSD latency has dropped >100× and GPU/AI‑chip compute throughput has risen >100×, making network latency the dominant component of end‑to‑end delay (≈60 %).

Remote Direct Memory Access (RDMA) reduces intra‑server transfer latency to ~1 µs and offloads the CPU, but it requires a lossless Ethernet substrate. Traditional IP Ethernet loses throughput when packet‑loss rates exceed 10⁻⁵, and mechanisms such as PFC/ECN only throttle traffic instead of preserving throughput.

Distributed architectures generate incast traffic spikes and large‑packet flows, causing severe congestion and packet loss.

Core Metrics for a Hyper‑Converged Network

The next‑generation fabric must simultaneously achieve:

Zero packet loss – essential for RDMA throughput.

Ultra‑low latency – sub‑microsecond per‑hop latency to keep network delay comparable to storage and compute.

High throughput – 25 Gbps/100 Gbps/400 Gbps links to satisfy AI data volumes.

These metrics depend on congestion‑control algorithms. Conventional DCQCN requires configuring dozens of parameters per NIC, leading to a combinatorial explosion of settings and sub‑optimal performance across heterogeneous traffic patterns.

Hyper‑Converged Network Architecture

The solution builds a lossless Ethernet fabric using RoCEv2 (RDMA over Converged Ethernet) together with Huawei’s iLossless intelligent lossless algorithm. Three technology pillars cooperate to eliminate congestion‑induced loss:

Traffic‑Control : End‑to‑end rate limiting, PFC deadlock detection, and proactive deadlock prevention.

Congestion‑Control : Global congestion management with AI‑enhanced ECN, iQCN, ECN Overlay, and NPCC, addressing the limitations of traditional DCQCN.

Intelligent Lossless Storage Network : iNOF (Intelligent No‑Loss Fabric) provides rapid host‑side control for storage traffic.

Key Technical Details

RoCEv2 runs over a lossless Ethernet substrate; losslessness is enforced by iLossless, which dynamically adjusts PFC and ECN thresholds.

AI‑based ECN generation predicts congestion before buffer overflow, allowing early traffic throttling without packet drops.

iQCN (Intelligent Quantized Congestion Notification) reduces the granularity of congestion signals, improving responsiveness for short‑lived flows common in AI training.

NPCC (Network‑wide Packet‑level Congestion Control) coordinates congestion feedback across all switches, ensuring consistent behavior in large‑scale topologies.

Comparison with Traditional Hyper‑Converged Infrastructure (HCI)

HCI integrates compute, storage, networking, and virtualization into a single appliance, requiring re‑architecting of existing resources. The hyper‑converged network focuses solely on the networking layer, delivering a unified Ethernet‑based fabric that can be deployed without modifying compute or storage stacks, enabling low‑cost, rapid scaling.

Performance Evidence

Independent testing (EANTC) shows up to 44.3 % latency reduction in HPC workloads and a 25 % increase in IOPS for distributed storage, while guaranteeing zero packet loss.

Network CAPEX typically accounts for ~10 % of data‑center investment, offering a 10× leverage effect compared to compute/storage costs.

SDN‑enabled lifecycle automation reduces OPEX by >60 %.

References

Huawei Support Encyclopedia: https://support.huawei.com/info-finder/encyclopedia/zh/index.html