Six Proven Methods to Optimize Server Capacity and Cut Costs in Large‑Scale Social Networks

SNG, Tencent's Social Network Operations division, manages nearly 100,000 Linux servers to support massive services such as QQ (2.47 billion daily active users) and QQ Space (5.96 billion monthly active users). To sustain growth while controlling operating costs, the team devised a refined capacity‑management approach that has saved the company over a hundred million yuan each year for two consecutive years.

1. Performance Management Method

CPU utilization is the primary metric for server efficiency. Uneven load across multi‑core CPUs can inflate costs. The team introduced a "CPU range" metric:

<code>CPU(range) = CPU(max) - CPU(min)</code>

If the CPU range exceeds 30 %, the device is flagged for optimization (e.g., multi‑queue NIC tuning and CPU affinity). A similar "module CPU range" metric is applied across distributed clusters:

<code>Module CPU range = CPU of highest‑load IP - CPU of lowest‑load device</code>

A module with a CPU range over 30 % indicates inconsistent capacity and requires remediation.

2. Density Management Method

Memory usage is better measured by "access density" rather than raw utilization. The formula is:

<code>Access Density = Packet Volume / Memory Used</code>

Consistent memory access density across devices within a module signals balanced load; deviations trigger corrective actions. This method also applies to SSD usage.

3. Feature Management Method

Analogous to QPS monitoring, this method evaluates whether business logic performance is optimal under specific scenarios. For example, long‑connection modules (QQ, QQ Space, Xinge) can be compared by the number of long connections per GB of memory, highlighting modules that need performance tuning.

4. Fragmentation Management Method

Small‑traffic clusters often waste resources when deployed as physical machines. By leveraging virtualization to fragment hardware resources, these clusters achieve both cost efficiency and high availability. Tencent's PaaS "Hive" platform, built on the SPP framework, further addresses capacity challenges for tiny services.

5. Barrel (Wooden‑Bucket) Management Method

Platform‑level services (QQ, QQ Space, QQ Music) employ a three‑site active‑active disaster‑recovery architecture (SET). Capacity is quantified per SET based on metrics such as concurrent users and core request volume. The overall capacity follows the "shortest‑board" principle: the SET’s maximum capacity is limited by its weakest module.

By forecasting stable concurrent‑user numbers, the required number of SETs can be pre‑planned, enabling cost‑effective multi‑site deployment.

6. Hardware Selection Method

Addressing hardware bottlenecks reduces per‑machine operating costs. Upgrading from 2 TB to larger‑capacity disks (4 TB, 8 TB) lowers storage cost per unit. In compute‑intensive scenarios (e.g., facial recognition, content moderation), replacing CPUs with GPUs yields significant performance gains for UGC storage workloads.

These six capacity‑management practices enable sustainable growth for social‑UGC services where user data continuously expands.