Repository navigation
AKO-154 Export offline PRISM-time replanning and native simulation comparison #154
Description
Generate a new experimental PRISM-time placement using evidenced materialization model, then export the exact same execution request for native SimGrid. No real submission, provisioning or daemon restart. Preserve frozen forecasts and identify mixed historical calibration provenance and incomplete route coverage. Local-only implementation and tests.
Activity
ovvesley commentedon Oct 11, 2026
Local implementation and experimental run complete. Full input scope: 14 schedulable resources (10 micro, 2 medium, 2 HPC), measured compute profiles, latest two evidenced transfer classes, shared fitted source capacity and bulk slots. PRISM-time defaults 120 beam/25 candidates: first candidate predicts 2321.672536 seconds. Native SimGrid exact selected request: 58/58 completed, 2368.118320 seconds (+2.00% vs planner, -16.49% vs previous fixed-placement simulation). Placement: micro01 8, micro02 9, micro03 8, Bora001 33. Physical network 4150785071 bytes vs prior 4383304054. Zero missing/duplicate tasks, micro overlaps or transfer-before-consumer barrier violations. Race tests passed for exporter/planner/provider. No real execution, cloud provisioning, daemon restart or database access. Raw confidence high is NOT empirical model validation; auxiliary service calibration and missing route coverage remain limitations. Plan and simulation costs use different billing bases. Safe report: experiments/montage-micro-hpc-e2/prism-time-transfer-experimental-20261011.json. Local branch AKO-154/offline-prism-replan, no push/PR.
ovvesley commentedon Oct 11, 2026
User authorized independent real execution. The 39.5-minute offline plan was correctly blocked at preflight: live PlaFRIM user quota available 20091539456 bytes vs conservative reservation 20234895697. No run or VMs launched for that failed preflight; experiment workspace directory confirmed empty. Replanned with shared live quota snapshots (same catalog/transfer model), yielding 10 micro and 18 HPC tasks across Bora/Diablo. New frozen prediction: PRISM 3127.922241s, native SimGrid 3163.497679s, 58/58 simulated. Registered plan via API and started montage-prism-transfer-real-20261011-v2 at 17:17:35 UTC. Daemon status running and Slurm first job 4921834. GCP quota/IP/disk and immutable image READY verified; target relay SSH source rules refreshed to current /32 and cpuCapacity1/speed0.25 restored. No direct database access, no daemon restart, no forecast recalibration during real. Safe forecast report experiments/montage-micro-hpc-e2/prism-time-quota-aware-real-20261011.json. Actual results pending; do not mark validated or complete.
ovvesley commentedon Oct 14, 2026
Offline PRISM-time replanning and native simulation use the same exported frozen request. Experimental comparison/provenance is preserved; agreement is not independent real validation. Located through AKO issue-number commits 557f825,99a1879 and verified integrated in published main 4aecc4c via PR #175. Full Go suite was rerun on main successfully for this audit; PR Go/docs CI is green. Closure covers this implementation/investigation scope only; remaining scientific forecast gates are tracked separately.