Breaking Barriers in Atomic Scaling: A Hardware-Software-Collaborated Framework to Deconstruct RDMA Atomic

Remote Direct Memory Access (RDMA) Atomics are widely adopted to ensure correctness in distributed synchronization. However, their scalability is still seriously constrained by internal locking within RNICs. This paper provides systematic analysis that uncovers fundamental bottlenecks when scaling atomic operations. To break the constraints in RDMA Atomic scaling, we present Fusa, a framework that transparently coordinates server-side hardware (RNIC) and software (CPU) to accelerate atomic executions. Fusa integrates fine-grained contention detection with selective onloading, which executes uncontended operations in hardware while redirecting contended ones to software. Fusa further designs a consensus mechanism for strategy switching. We evaluate Fusa using microbenchmarks and unmodified RDMA-based systems, showing that Fusa improves throughput by up to 4.6×.