| name | habu-host-profiles |
| description | Use when timing the Habu native test suite on a supported host class, comparing host profiles, or setting timing-regression policy. |
Habu Host Profiles
The test suite auto-detects its host profile by default. Use
--perf-profile NAME only to force a specific profile. Profiles set pool slots,
nested pool slots, and timing budgets. Put manual overrides after
--perf-profile.
Profiles:
macos-arm64-10x2: macOS ARM64 target, --pool-slots 10,
--nested-pool-slots 2, hot and cold --budget-ms 30000,
--wall-budget-ms 35000.jetson-orin-clocks-4x2: Linux target on NVIDIA Jetson with CPUs 0-7
online, --pool-slots 4, --nested-pool-slots 2, hot
--budget-ms 100000, --wall-budget-ms 110000; cold 150000 / 160000.linux-arm64-4x2: generic Linux ARM64 target, --pool-slots 4,
--nested-pool-slots 2, hot --budget-ms 120000; cold 150000.dgx-spark-10x2: Linux target on an NVIDIA DGX Spark (GB10), --pool-slots 10,
--nested-pool-slots 2, hot and cold --budget-ms 30000,
--wall-budget-ms 35000. Detected from the SMBIOS/DMI product family
(/sys/class/dmi/id/product_family contains DGX Spark), since the GB10
exposes no /proc/device-tree/model.
The runner's --wall-budget-ms uses its monotonic elapsed test-suite time;
wrap the command with /usr/bin/time -p when exact process wall time matters.
Top-level --pool-slots is capped at 12.
macOS hot profile:
/usr/bin/time -p bin/hb --load test/run.f -- --under bin/hb --timings
macOS cache-fill profile:
/usr/bin/time -p bin/hb --load test/run.f -- --under bin/hb --cold-cache --timings
Jetson/Orin prep:
sudo jetson_clocks --store /tmp/habu-clocks-before.conf || true
for c in 4 5 6 7; do echo 1 | sudo tee /sys/devices/system/cpu/cpu$c/online >/dev/null; done
sudo jetson_clocks
Jetson/Orin hot profile:
/usr/bin/time -p bin/hb --load test/run.f -- --under bin/hb --timings
Jetson/Orin cache-fill profile:
/usr/bin/time -p bin/hb --load test/run.f -- --under bin/hb --cold-cache --timings
DGX Spark (GB10) hot profile (auto-detected):
/usr/bin/time -p bin/hb --load test/run.f -- --under bin/hb --timings
DGX Spark (GB10) cache-fill profile:
/usr/bin/time -p bin/hb --load test/run.f -- --under bin/hb --cold-cache --timings
DGX Spark calibration note: the GB10 has heterogeneous cores — 10 Cortex-X925
performance cores at 3.9GHz (the fixed calibration spin measures 87ms) and 10
Cortex-A725 efficiency cores at 2.8GHz (132ms). The perf verdict brackets each
attempt with a pre-run and a post-run calibration spin and invalidates the
attempt on >10% drift. The spin has zero in-process drift, but the long-lived
gate driver launches on a performance core (pre-cal 87ms) and, while it blocks
waiting on the pool, is migrated onto an efficiency core; its single in-process
post-cal spin then starts there and migrates back mid-spin, reading a blended
~104ms and tripping the drift gate. The runner therefore measures the post-run
calibration in a freshly spawned child (test/cal-spin.f), which the scheduler
places on a performance core again, so pre and post both read ~87ms and the
verdict is admissible. The committed reference is TR-CAL-REF-SPARK-MS = 87.
Do not "fix" a spark drift red by widening the stability threshold.
--cold-cache uses a private per-run cache root and the profile cold budget. It
does not delete the persistent content cache.
Default persistent-cache runs also mark themselves cold if a source change
invalidates HABU_UNDER_TEST or builder cache artifacts during setup.
