Hermes: Enhancing Layer-7 Cloud Load Balancers with Userspace-Directed I/O Event Notification#
SIGCOMM 2025 digest Program Core pick Coimbra · 8–11 September 2025
Tian Pan et al. (Alibaba Cloud)
Why: eBPF on the datapath solving a real L7 LB connection-distribution problem (epoll-exclusive LIFO skew, reuseport blindness to worker health) is squarely cloud-native networking.
What: A framework making userspace worker status drive kernel connection dispatch, deployed on O(100K) cores across 33 Alibaba regions, cutting daily worker hangs 99.8% and L7 LB unit cost 18.9%.
How: Lock-free worker-status management plus eBPF overriding reuseport socket selection in a closed loop from userspace to kernel.
Where applicable: General Linux/eBPF technique usable in any Envoy/proxy-style fleet; the scale numbers are Alibaba-specific.
