Tencent AMS Cloud Migration: Architecture, Solution Selection, and Optimization Process

The article describes the 2020 cloud‑migration project of Tencent’s Advertising Marketing Service (AMS) line, whose theme was “grow on the cloud, reduce cost and increase efficiency”. By leveraging Tencent Cloud technologies and a self‑built migration framework, the team achieved significant cost reduction and performance improvement.

1. Business Architecture – The core advertising playback chain handles billions of daily requests, with a complex topology of over 40 modules that must respond within 100‑200 ms. High request volume, strict latency, and massive data traffic put extreme pressure on bandwidth and storage.

2. Cloud‑Migration Solution Selection – Two main options were considered: CVM (cloud virtual machines) and TKE (Tencent Kubernetes Engine). Because most AMS modules require >80‑core machines, CVM was chosen as the primary platform, with TKE as a secondary option for future expansion.

3. Migration Phases – The migration was divided into three stages: (1) Infrastructure migration – moving IDC traffic to cloud VMs and trialing cloud databases and load‑balancers; (2) Architecture upgrade – extensive use of PaaS services such as auto‑scaling and Docker; (3) Cloud‑native R&D‑Ops – building a cloud‑centric development and operation ecosystem.

4. Optimization Process

• Baseline environment adaptation – initial testing revealed performance gaps between cloud VMs and physical machines.

• CVM performance tuning – measured latency using percentile statistics; identified an 8‑10 ms increase in 99‑percentile latency for cloud VMs compared to physical machines.

• Ping jitter issue – discovered that outdated ethtool on the OS caused network‑queue problems; resolved by upgrading the tool and the kernel.

• CPU jitter and avalanche – high CPU usage and cross‑core scheduling caused cache misses; mitigated by isolating CPU cores with DPDK and adjusting scheduling policies.

• AMD CPU optimization – AMD instances showed severe jitter under load; a single‑node AMD VM with 90 logical cores and dedicated socket reduced latency by >25 % and cost to 60 % of the Intel baseline.

• CBS storage optimization – bandwidth limits of cloud block storage forced a redesign of data transfer: switching to incremental files, using /dev/shm for in‑memory processing, and adding primary/backup flags to halve write volume. Resulted in I/O reduction from >100 MB/s to ~30 MB/s.

• Network latency reduction – deployed PVGW devices and direct‑connect optimization, cutting ping latency between cloud VMs and IDC to 4‑5 ms.

• TKE service construction – evaluated STKE (internal) and TKEstack (private‑cloud) platforms; selected TKEstack for its rich network modes, elastic resource limits, GPU virtualization, and built‑in P2P image distribution. Implemented container‑based CI/CD pipelines using CSIGHUB hooks and BlueShield pipelines.

• Fault‑tolerance and disaster recovery – designed multi‑IDC deployment with redundant cloud and physical sets, ensuring seamless traffic switching and reducing average latency by 26‑35 %.

5. Benefits – The migration delivered >25 % latency improvement, cost reduction to 60 % of the original, higher stability, and a scalable cloud‑native architecture that supports future growth of AMS services.