Two notebooks for running NVIDIA BlueField-3 DPUs on FABRIC, meant to be worked through in order.

1. Create a slice using BlueField DPUs

bluefield_dpus_l2_network_basic_auto.ipynb — Creates a slice with a BlueField-3 SmartNIC, installs the BFB bootstream on the DPU, brings up the rshim/tmfifo management path to the ARM cores, and configures NAT so the DPU can reach the internet. Ends with a working network and shell access to the DPU.

2. OVS hardware offload

bluefield_dpu_ovs_hardware_offload.ipynb — Demonstrates the actual reason to use a DPU rather than a fast NIC. Open vSwitch runs on the DPU's ARM coresto the wire, so every packet crosses the DPU; the question is whether it is s ARM or in hardware by the ConnectX-7 eSwitch. The notebook builds an expli (p0 ↔ pf0hpf), then runs the same hw-offload on and off, comparing throughput, TCP retion, and how many flows land in hardware.

Offload relieves the DPU's ARM cores, not the host's vCPUhe ARM over the rshim link.

Contents

• bluefield_dpus_l2_network_basic_auto.ipynb —
• bluefield_dpu_ovs_hardware_offload.ipynb — of
• node_tools/bf3_rshim.sh — host-side forwardin (--mode ipv4 | ipv6 | dual). Keep this direct

The offload example needs a BlueField on one node only; the p NIC_Basic, which is a shared vNIC and caps absoluteul results are the delta between runs and the ARM CPU cost, not peak Gbps.<

Source: fabric-testbed/jupyter-examples under fabric_examples/fablib_api/bluefields/