Dynamic Load Balancing for Direct Server Return Networks Using eBPF for In-Band Metric Feedback

Direct Server Return (DSR) enables backend servers to send responses directly to clients, bypassing the load balancer on the return path. Removing that extra hop trims end-to-end latency and prevents the balancer from becoming a bottleneck at high request rates. This paper introduces a backward-compatible DSR variant that encodes each server’s load metric inside an Internet Protocol (IP) option, so the metric travels with ordinary data packets, and no polling traffic is needed. A Linux extended Berkeley Packet Filter (BPF) prototype adds only a small patch to the data path, yet yields up to 47% more requests per second than an explicit-polling baseline, requiring no changes to either the client or server. The proposed solution does not modify application logic and supports dynamic load balancing in heterogeneous and variable workloads, such as microservices, batch processing, or machine learning inference. It is fully deployable on commodity servers, runs entirely in kernel space, and eliminates separate metric-collection traffic. Performance evaluation demonstrates significant throughput and latency improvements needed for large-scale and low-overhead load balancing of real deployments.