NCCLbpf: Verified, Composable Policy Execution for GPU Collective Communication

This collection covers recent systems research with direct operational and security relevance: NCCLbpf embeds a verified userspace eBPF runtime into NCCL to run composable, verifiable GPU communication plugins with atomic hot-reloads and very low latency overhead, reducing unsafe plugin behaviors at load time; Carnot protocols show how erasure-coding and multi-round finality can minimize leader bandwidth in Byzantine-tolerant SMR; a decentralized multi-domain orchestration architecture proposes SDN-enabled anomaly detection (FU-HST) to harden decentralized federated learning against Byzantine/