Keep a benchmark record another engineer can reproduce
Download a benchmark manifest validator and HPC or AI acceptance worksheet that retain workload, placement, cache conditions and evidence.
Write down the question before the command
A run record connects a result to its workload, environment and acceptance criterion. The template separates input identity, software versions, topology, placement, cache policy, warmup, duration, repetitions, metric and evidence references.
Download benchmark-manifest.json · Download validate-manifest.py · Download acceptance-checklist.csv
python3 validate-manifest.py benchmark-manifest.json
The supplied template has status not_run and value not_measured. It is intentionally illustrative. Replace all placeholders before collecting evidence; its example durations and repetitions are not a recommendation for your benchmark.
What validation does and does not establish
The validator checks required sections, selected nested fields, numeric ranges and consistent result states. A pass or fail entry requires a finite numeric value and evidence references. A not_run entry requires not_measured. The validator does not run the workload, inspect referenced files, prove the identity of an input, approve a criterion or attest that a claimed pass is true.
A structurally valid template is a starting record. Review it with the engineers and researchers who own the workload. Retain actual logs, immutable container identities where used, topology and selected counters alongside the record.
Agree on completion
Use the checklist to record each criterion’s owner, evidence and status. Include correct output, concurrent demand, storage behavior, checkpoint recovery, configuration persistence and administrator handover. Keep unresolved items visible in the final report.
See a sample acceptance report · Follow the cluster-acceptance guide · Read about acceptance testing services.