Migrating from SPserver to BRPC: Performance Optimization and Lessons Learned

Performance optimization is a critical issue for backend services, especially in advertising where timeouts affect both client budgets and end‑user experience; the article begins by describing the SPserver framework, its historical advantages, and the emerging problems such as memory consumption, concurrency limits, and frequent timeouts.

Through detailed measurements and reproducing timeout scenarios, the authors identify that socket buffer queues, uneven thread CPU utilization, and the single‑threaded reactor model of SPserver cause bottlenecks, especially when a single thread handles all I/O for many sockets.

Root‑cause analysis reveals that SPserver’s single‑threaded I/O design cannot scale under high‑concurrency workloads, prompting the decision to replace it with Baidu’s BRPC framework.

BRPC is chosen for its high‑concurrency, high‑performance characteristics, rich documentation, and proven production use; the article outlines BRPC’s M:N threading model (bthreads), its ability to handle separate send/receive threads per socket, and the resulting elimination of the “one‑place‑block‑everywhere” problem.

Performance tests show that BRPC fully utilizes multi‑core CPUs, reduces context‑switch overhead by keeping most switches in user space, and achieves near‑zero socket queue lengths, leading to a QPS increase from ~50k to ~100k and a halving of service instances.

The authors also share practical tips: handling thread‑local storage when using bthreads, cautious use of CPU profilers, considerations when comparing BRPC with gRPC (favoring multi‑channel communication), and choosing appropriate bthread launch APIs (background vs. urgent) based on workload.

In conclusion, migrating from SPserver to BRPC yields significant performance gains, simplifies concurrency management, and provides useful built‑in monitoring tools for further optimization.