- Migration of Rust WebRTC SFU from Tokio's work-stealing model to thread-per-core (TPC) architecture reduced P99.99 latency from 70ms to 10ms and increased capacity by 25%.
- TPC architecture eliminates unnecessary async channels, processing packets synchronously to avoid scheduling overhead and priority inversion, improving tail latency.
- Benchmarks show stable latency at high CPU loads, with TPC maintaining low jitter even at 80% CPU utilization, enhancing congestion control and stream quality.