CoPilotIO: CPU as a Co-Pilot for GPU I/O to Free GPU Compute

Limited GPU memory increasingly forces modern AI and data analytics workloads to access terabyte-scale datasets and model states from storage, making efficient GPU I/O critical. Existing GPU I/O engines are either CPU-centric or GPU-centric. CPU-centric approaches avoid consuming GPU resources but often fail to provide high-throughput, ondemand GPU access due to kernel overheads and limited parallelism. GPU-centric approaches enable fine-grained ondemand I/O but require intensive I/O polling that consumes valuable GPU resources and introduces intra-warp, inter-warp, and inter-SM I/O stalls. We present CoPilotIO 1 , a novel GPU I/O engine that delivers high-throughput, on-demand storage access without sacrificing GPU compute resources. CoPilo-tIO adopts an asynchronous GPU I/O architecture in which GPUs initiate I/O while CPU cores act as I/O co-pilots responsible for completion polling. To enable efficient coordination, CoPilotIO introduces a split SQ/CQ architecture, hardware barrier-based synchronization, a lock-free barrier-table, and adaptive CPU-GPU co-polling. Across microbenchmarks and real applications, including GoFS, LLM Mixture-of-Experts (MoE) inference, and Deep Learning Recommendation Models (DLRM), CoPilotIO reduces I/O-induced stalls by up to 55.5%, requires 50% fewer SMs to saturate the GPU PCIe bandwidth, accelerates GoFS by up to 17.4%, and improves application performance by up to 85%.