Toward eBPF-Accelerated Pub-Sub Systems

Publish-subscribe (pub-sub) systems are a fundamental building block for real-time distributed applications, where high throughput and low latency are critical. Existing brokers can suffer performance bottlenecks as they operate in user space and rely on the socket API and full kernel stack traversal for every message. We present BPF-Broker, a novel pub-sub broker that leverages eBPF to accelerate message dissemination by decoupling the broker's control and data paths. Subscriber management is handled in user space, while message forwarding is done early in the kernel using the TC ingress and XDP hooks. Our evaluation shows that BPF-Broker achieves up to 3× higher throughput compared to our Socket-based baseline broker under high subscriber counts, and up to 2-10× lower end-to-end latency. These results highlight the potential of eBPF in accelerating pub-sub systems.