Workshops Archive

DPUCrumbs: Dissecting DPU Pathways


Workshop: PDSW’26: The 11th International Parallel Data Systems Workshop

Authors: Ray A. O. Sinurat (University of Chicago), Sarah Neuwirth and Radita Liem (Johannes Gutenberg University Mainz), Jay Lofstead (Sandia National Laboratories), Nikoli Dryden (Lawrence Livermore National Laboratory), and Hariharan Devarajan (Lawrence Livermore National Laboratory (LLNL))

Abstract: Data processing units (DPUs) such as NVIDIA's BlueField-3 increasingly used to offload MPI collectives, compression and data movement in HPC. Yet cost of offload is spread across every layer between the application and the card. Existing tools observe one layer at a time, each on its own clock, so host, driver and card cannot be joined into one timeline. A slow offload is therefore observable but not explainable. We present DPUCrumbs, an eBPF-based tracer that extends DataCrumbs across PCIe onto the card. It captures application calls, DOCA and ibverbs entry points, MLX5 driver and kernel's DMA and IOMMU paths. NIC hardware timestamps then place both endpoints on one time axis. On a DOCA RDMA ping-pong, DPUCrumbs measures PCIe crossings at 13.2 and 8.6 µs, shows the DPU responding in 0.82 µs and and and reveals that reply waits 6.1 µs at the NIC before host consumes it, at 3.6% overhead.


Back to PDSW’26: The 11th International Parallel Data Systems Workshop Archive Listing

Back to Full Workshop Archive Listing