GDSight: Per-Operation Cross-Layer Attribution for GPUDirect Storage
Authors: I. Yildirim, X.-H. Sun, A. Kougkas
Date: November, 2026
Venue: The 11th International Parallel Data Systems Workshop (PDSW'26), held in conjunction with SC26, Chicago, IL, USA
Type: Workshop
Abstract
GPUDirect Storage (GDS) copies NVMe data into GPU memory, bypassing the CPU data path and the identity linking an operation to the device work it causes. To our knowledge, no external GDS tool attributes a device command to its causing application operation: cuFile stops at the call, counters report only aggregates, and time-based matching fails once submission is asynchronous, even at one operation in flight. We present GDSight, which attributes each NVMe command to the operation that caused it, whether it entered cuFile or fell back to POSIX. Using eBPF, it matches commands by timing and by physical address, and only address survives asynchrony. On unmodified KvikIO, NIXL and HDF5 we pin pathologies to their cause: a silent bypass cuFile’s counters miss, 4× device amplification, and 22× tail inflation. We show that each diagnosis resolves to a measured intervention, worth up to 17× where throughput is the limit, or none.