WorksComputer Science
GNoF: A Low-SM-Overhead GPU-Native NVMe-over-Fabrics Completion Path for Data-Intensive GPU Workloads
Abstract
GPUs are the main compute engine for data-intensive workloads whose datasets exceed device memory and reside on remote flash. The conventional host CPU orchestrates every NVMe-over-Fabrics (NVMe-oF) transaction and stages data through host memory, costing high CPU usage, repeated CPU-GPU synchronization, and redundant data movement. GPU-initiated storage removes the CPU from the I/O path, but existing designs leave two problems: synchronous lock-step completion pins streaming multiprocessors (SMs) in spin-polling, wasteful over a 10-100 microsecond RDMA round trip, and current GPU-native remote systems focus on target-side firmware while using per-command doorbells and periodic CQ polling. We present GNoF, a GPU-native NVMe-oF/RDMA data path where GPU threads build command capsules, post send work-queue elements, ring doorbells, and reap completions entirely on the device; the CPU only handles one-time connection setup, the admin queue, and a keep-alive thread. Three mechanisms target long fabric round trips: a doorbell mover that coalesces a burst's doorbells into one MMIO ring; a batched CQ poller with a lock-free fast peek and monotonic completion counters that cannot lose a wakeup; and a set-associative, reference-pinned GPU page cache with split fetch/wait pipelining and orphan-fill takeover. GNoF needs no target or firmware change and runs against an unmodified SPDK target. On a 200 Gb/s InfiniBand fabric, GNoF reaches 1.77 million 4 KB IOPS per client at 0.37% host CPU and about 1% SM occupancy, versus 40% CPU for kernel nvme-rdma and about 100% SM occupancy for lock-step polling. Against three existing baselines (kernel nvme-rdma, BaM over a BlueField-3 DPU, and BaM on a local SSD) under identical workloads, we show 10-22x gains on selective column I/O, CPU-burn immunity, and the regime where bulk staging is faster.