Tencent RegionEIP: High‑Performance Networking with X86 & P4

Overview of Tencent RegionEIP

Tencent Mountain Sea Gateway (RegionEIP) integrates core products such as Cloud Load Balancer (CLB) and Elastic IP (EIP) to provide multi‑operator public network access for Cloud Virtual Machines (CVM), CLB and other services. It emphasizes high reliability, strong scalability, high performance and robust DDoS resistance.

X86 Network Architecture

In the X86 architecture, each server is called a Load Distributor (LD). An X86 RegionEIP cluster spans two zones, with four LDs in each zone. All LDs share identical forwarding rules, so traffic can be forwarded correctly regardless of which LD it reaches.

Key high‑availability mechanisms include:

Zone disaster recovery : Different IP segments are announced in each zone; if the primary set of LDs fails, traffic switches to the backup set within 3‑5 seconds.

Dual‑link redundancy : Each LD has two external and two internal ports; if a port fails, the other takes over without traffic loss.

Power redundancy : LDs are distributed across separate power supplies, ensuring continuity when a single power source fails.

Device redundancy : When two ports on an LD fail, the control program isolates the device and reroutes traffic via other LDs.

Traffic shaping uses a DPDK‑based “distributed feedback algorithm” that samples, calculates, and distributes quotas, but it suffers from latency and burst‑traffic limitations.

P4 Network Architecture

To overcome X86 limitations, RegionEIP adopts P4 programmable switches built on Tofino chips, which provide native rate‑limiting and high‑precision traffic control.

The P4 cluster also spans two zones with eight switches (four per zone). Each switch connects to four WAN‑core (WC) and four LAN‑core (LC) devices via 2 × 100 Gbps links, delivering 800 Gbps inbound and outbound capacity per switch. The cluster can horizontally scale up to eight switches.

Similar to X86, the P4 cluster uses zone disaster recovery and a four‑level routing priority based on as‑path attributes to ensure that a given IP segment is processed by only one switch.

Four‑level routing priority example (simplified):

Level 1: highest priority

Level 2: second priority

Level 3: third priority

Level 4: lowest priority

If the highest‑priority device fails, traffic is automatically taken over by the next level.

Port Redundancy Schemes

Two designs are evaluated for port‑level redundancy:

Scheme 1 – ECMP groups : WAN and LAN ports are grouped; hash‑based ECMP distributes traffic. Advantages: detection program does not need to know port pairing; a single port failure only affects its side. Disadvantages: possible traffic imbalance, slower convergence, and harder troubleshooting.

Scheme 2 – Paired ports : Each WAN port is paired with a specific LAN port. Advantages: deterministic traffic path, 1:1 convergence ratio, easier operation. Disadvantage: detection program must be aware of port pairs.

After comparison, Scheme 2 is chosen for its consistent convergence and operational simplicity. Additional thresholds trigger device isolation when multiple ports fail, and the detection system operates at millisecond granularity to ensure rapid failover.

Conclusion

Both the X86 and P4 architectures provide strong forwarding capability, high stability, and cost‑effective solutions. The P4 design further improves resource utilization through pipeline folding and reduces costs by directly connecting to core switches. From an operational perspective, the designs are transparent to users, while offering enhanced disaster recovery, routing priority, and port‑level resilience.