Measures host and DPU-CPU control-plane rule-install rates of DOCA Flow pipelines.
npx skills add https://github.com/NVIDIA/skills --skill doca-flow-perf
Please help me install this Agent Skill. Skill: doca-flow-perf Source: https://github.com/NVIDIA/skills/tree/main/skills/doca-flow-perf Command: npx skills add https://github.com/NVIDIA/skills --skill doca-flow-perf
Getting a defensible DOCA Flow rule-install or rule-delete rate number is hard: results vary with the JSON policy, backend choice, worker configuration, DOCA version, and firmware, and naive comparisons against published reference numbers are misleading. This Skill guides agents through running the doca_flow_perf synthetic benchmark correctly so the resulting Kops/sec number is reproducible and interpretable.
Ask your agent to measure the host-side DOCA Flow rule-insert rate on your BlueField using doca_flow_perf with a 5-tuple hairpin policy from the shipped configs directory.
High-intent search queries and answers about installing and using this skill.
Use doca_flow_perf with a JSON policy from the shipped configs/ directory that matches your traffic class. Run a single-iteration smoke test at a small entry count first, then scale up workers and iterations while capturing the DOCA version, firmware, policy, and backend alongside the result.
doca-flow-perf measures the host or DPU-CPU control-plane rule-programming rate with a synthetic pipeline. doca-flow-dpa-perf measures the DPA-offloaded path, and doca-flow-tune optimizes an already-deployed live Flow application rather than measuring a synthetic one.
No. doca_flow_perf measures only the control-plane cost of installing, deleting, and optionally querying Flow rules in CPU cycles per iteration. It does not send packets through the pipeline, so dataplane throughput and end-to-end latency are out of scope.
Both backends share the same JSON policy contract but exercise different programming paths: the DPDK backend uses rte_flow while the DOCA backend uses the native doca-flow library. You must report which backend produced any number, since results are not comparable across backends.
Published numbers assume a specific DOCA version, BlueField generation, firmware, backend, and worker/queue/burst configuration. A comparison is only valid when that entire four-tuple matches; also check for high iteration variance, num_failed > 0, and host instability such as unpinned worker CPUs.
Yes, by copying the closest shipped configs/ exemplar and editing documented fields such as num_inserted_entries, val, and steps. Never invent JSON keys not present in the exemplars or the public DOCA Flow Perf guide, and re-run the smoke test after every edit.
NVIDIAConfigure and troubleshoot NVIDIA networking deployments on Kubernetes.
stolyarchukGuide C++ engineers in designing and optimizing DOCA Flow pipelines on NVIDIA SmartNICs.
openeuler-mirrorDiagnose NPU hardware, PCIe link, and driver failures from server logs.
gtgspotApply the DPA playbook to determine processor vs controller and generate a review memo.
