RDMA Performance Tester

Orchestrated RDMA benchmark: ib_write_bw/ib_read_bw between all host pairs, latency heat map, bandwidth bottleneck identification and comparison vs theoretical line rate.

NVIDIA

Linux RDMA

perftest

Open in Generator Open in Platform

You're viewing a preview

The full script, parameters, and execution console are available with a free account.

Already have access? Sign in →

Capabilities

Full-mesh ib_write_bw/ib_read_bw
Latency heat map
Line-rate efficiency %
Bottleneck pair identification
Per-NIC aggregate throughput
CSV + HTML export

Required inputs

Parameters the script accepts. Defaults shown; some are vendor- or context-gated.

Parameter	Type	Default	Notes
NIC line rate (Gb/s) line_rate_gbps	number	400	—
Min efficiency (%) efficiency_floor	number	90	—
Message size (bytes) msg_size	number	65536	—

Hint 1 — Name your vendors and OS versions

Mention the exact platforms you run (NVIDIA, Linux RDMA, perftest) so the generated ai/gpu dc logic uses the right CLI/NETCONF syntax.

Hint 2 — State your real thresholds

Provide concrete values for NIC line rate (Gb/s), Min efficiency (%), Message size (bytes) instead of the defaults — they shape what counts as a fault.

Hint 3 — Prioritize the checks you need

This tool can full-mesh ib_write_bw/ib_read_bw and latency heat map. Ask for the subset relevant to your incident to keep output focused.

Hint 4 — Describe your inventory format

Tell the assistant whether your device list is CSV, YAML, or JSON and which columns it has, so parsing matches your data.

Sample inventory schema

Authoritative shape of the device/policy data this script consumes.

No bundled inventory sample. The script accepts standard device lists (CSV/YAML) with hostname, mgmt IP, vendor, and credentials.

Expected output

Reference terminal output the script should produce — used as a stylistic and structural target.

terminal

[2026-06-12] INFO  Running full-mesh RDMA benchmark across 8 hosts (28 pairs)...

PAIR              BW(Gb/s)  EFF%   LAT(us)  STATUS
gpu01→gpu02       394       98.5   1.9      OK
gpu03→gpu05       331       82.8   3.4      DEGRADED ←
gpu06→gpu08       391       97.8   2.0      OK

[WARNING] gpu03→gpu05: 82.8% efficiency (<90%) — check path / congestion
Aggregate: 10.6 Tb/s · avg eff 95.1% · 1 degraded pair