Inside Vivo’s 2025 VDC: Traffic Management, Microservice Optimizations & AI GPU Platforms

01 Traffic Management Driving Growth

Vivo’s internet probing team explained how a global network of real smartphones serves as probes, providing massive, real‑time data that underpins intelligent traffic scheduling. By closing the loop of input, execution, processing, and application, manual, experience‑based dispatch was replaced with a data‑driven, proactive system that reduces operational labor while dramatically improving efficiency and reliability.

Building on this data foundation, the internet operations platform adopted a hybrid cloud architecture, unifying private and public clouds. Despite typical challenges—complex management, high costs, quality issues, low efficiency, and security risks—Vivo’s integrated traffic‑management platform achieved unified governance, smart decision‑making, security protection, and automated monitoring, with future focus on AI‑driven deep optimization and tighter coupling of network quality to business metrics.

02 Microservice Architecture & Dubbo Performance Optimization

To support rapid user growth, Vivo migrated to a microservice architecture, consolidating Dubbo as its Java RPC framework. For Dubbo routing, near‑site routing reduced latency for latency‑sensitive services, while simplifying routing chains and introducing bitmap caches boosted execution efficiency. In load balancing, Vivo enhanced the community P2C algorithm with weight calculations, enabling adaptive, smooth traffic distribution and achieving higher service quality and cost savings.

Future plans include aligning Dubbo with upstream open‑source versions and building a language‑agnostic microservice governance platform that offers unified service discovery, traffic monitoring, governance, and observability, thereby lowering system complexity and operational costs.

03 Full‑Link Multi‑Version Environment Management

Vivo’s DevOps team introduced a “full‑link multi‑version” strategy that automates environment provisioning from component dependencies to complete chain readiness, enabling parallel sandboxed versions that eliminate resource contention. The three key capabilities—full‑link readiness, parallel version isolation, and automated lifecycle management—deliver cost reduction, efficiency gains, and reliable, scalable development pipelines.

By orchestrating environments, elastic resources, and traffic isolation, Vivo achieved significant operational savings and plans to further standardize research environments and optimize resource costs through dual‑track strategies.

04 GPU Containers & AI Training Platforms

Vivo’s GPU platform, essential for large‑model AI workloads, consists of a physical layer, container platform layer, and AI engineering layer, providing a robust compute foundation. Custom virtualization enables “one card, three uses,” while elastic GPU scaling automatically adapts to load, reducing idle resources and operational overhead.

The VTraining platform, built atop the container capabilities, supports core products like Blue‑Heart Xiao‑V. By minimizing infrastructure failures and optimizing task handling, daily failure rates dropped from 2% to 0.1‰ and effective training time rose from 60% to 99%. GPU utilization improvements stem from low‑priority task scheduling, tidal deployment, and virtualization, achieving high‑efficiency resource reuse.

Future directions include multi‑cluster scheduling, mixed offline GPU workloads, and GPU resource pooling, as well as enhancing AI training stability and fine‑grained GPU resource operations.