GalaxyLink experiment · Stage: Simulation
Workload scheduling under changing power limits
How do simple scheduling rules change on-time work, restart effort and estimated cost under the same changing power limit?
Record date 2026-11-01 · By Eby Mathew
Scope and inputs
Offline Python model, synthetic workload and tariffs. No live control or measured GPU/site performance.
- Jobs
- 14
- Time slots
- 24
- Slot length
- 15 min
- Inputs
- Bundled synthetic.json
Results
SIMULATION ONLY. No real-world savings or performance claim.
| Policy | On-time jobs | Missed deadlines | Interruptions | Energy (kWh) | Est. cost (AUD) |
|---|---|---|---|---|---|
| First come, first served | 5 / 14 | 7 | 4 | 21.25 | $4.54 |
| Earliest deadline first | 3 / 14 | 9 | 7 | 21.25 | $4.51 |
| Power-aware packing | 1 / 14 | 11 | 19 | 22.00 | $4.66 |
Power packing did not establish an advantage. Repeated interruptions increased modelled restart effort. These inputs do not prove general policy superiority.
What this cannot support
- Synthetic workload, power trace and tariffs.
- Perfect preservation of completed work at slot boundaries.
- Constant active power and simplified fixed restart costs.
- No cooling, idle-server, network, storage, battery or physical microgrid dynamics.
- No measured customer savings, market validation or production reliability.
Validation: 10 behavioural checks passed (Python 3.12.14 on Linux). Checks model behaviour and accounting, not real-world performance.
Next experiment
Find an operator with a specific unmet need; agree an authorised dataset, competent baseline, realistic recovery assumptions, fair service metrics and a bounded feasibility study.
Sources
[1] GalaxyLink · Internal primary experiment
GalaxyLink offline workload and power replay experiment — version 0.1
Supports: The original synthetic experiment, its documented model, output figures and limited validation.
Reviewed 2026-11-01