Dynamic Hierarchical Birkhoff-von Neumann Decomposition for All-to-All GPU Communication

Two systems papers on GPU/cluster networking and RDMA fault tolerance. The first proposes a dynamic hierarchical Birkhoff–von Neumann decomposition plus dynamic frame sizing for two-tier GPU fabrics to reduce all-to-all completion time, mitigate intra-server GPU/NIC skew, and provide provable stability for online scheduling. The second (SHIFT) proves an impossibility Trilemma for cross‑NIC RDMA failover (Exactly‑Once Execution, Receiver‑NIC Opacity, and Zero‑Copy cannot all be preserved) and presents a user-space rdma-core implementation that achieves cross‑NIC failover by relying on idempotid