RTCBench: A Scalable Benchmarking Framework for Multi-Party Video Conferencing
Abstract
Video conferencing has become an essential medium for modern communication, serving billions of users worldwide in daily interactions. Despite this pervasiveness, the scalability of its underlying infrastructure has received limited attention from the research community, largely due to the lack of a standardized and reproducible benchmarking approach. Realistic evaluations require many participant clients, yet conventional browser-based emulation incurs heavy client-side media processing costs and often makes benchmark hosts themselves the CPU bottleneck before the conferencing systems become saturated. Our empirical analysis shows that video capture/rendering and encoding/decoding dominate endpoint CPU overhead, revealing a critical barrier to large-scale benchmarking. To address this challenge, we present RTCBench, a lightweight and extensible framework for stress-testing large-scale video conferencing systems. RTCBench employs lightweight sender and receiver designs that eliminate unnecessary runtime media processing while preserving the packet-level transport behavior, delivery statistics, and feedback loops needed for realistic conferencing dynamics. Its configuration-driven orchestration and plugin abstraction further support systematic evaluation across heterogeneous conferencing platforms. Our experiments demonstrate that RTCBench substantially outperforms Chromium-based endpoints, sustaining up to 86× more emulated participants on the same hardware and enabling scalable and reproducible benchmarking on modest resources.