What our first simulation actually taught us

The first GalaxyLink scheduling experiment produced a useful reminder: a more complicated rule does not automatically deliver a better outcome.

We compared three simple policies across the same synthetic workload and power trace. The example contained 14 jobs and 24 fifteen-minute slots. It included later-arriving urgent jobs and two periods of zero available capacity.

The results are model outputs, not measurements from a data centre.

What happened

Policy	On-time jobs	Late-completed jobs	Incomplete jobs
First-come-first-served	5	5	4
Earliest deadline	3	6	5
Power packing	1	6	7

The power-packing heuristic repeatedly interrupted work. Its modelled restart effort reached 19 slots, compared with three for first-come-first-served and eight for earliest deadline.

Those restart slots consumed modelled energy without advancing useful work.

This does not prove that first-come-first-served is generally superior. It shows what happened with these inputs, these rules and this particular interruption model.

Why the energy bill was insufficient

Earliest deadline recorded a slightly lower total simulated energy cost than first-come-first-served: A$4.51 against A$4.54. Yet it completed fewer jobs on time.

The tariffs were invented for the exercise. Even within that simplified setting, selecting a winner from the energy bill alone would hide a material difference in delivery.

Useful work, deadline performance and total cost need to be examined together. Where jobs differ in size or value, counting completed jobs also needs care.

What the checks establish

Ten behavioural checks passed using Python 3.12.14 on Linux. They covered issues including accounting, deadline boundaries, zero capacity, interruptions, invalid inputs and deterministic replay. Equivalent JSON and CSV inputs produced identical reports.

All three policies respected the simulated current power envelope.

These checks support confidence that the chosen model behaves as documented. They do not validate its representation of physical equipment, prove a commercial advantage or demonstrate reliability in production.

What remains missing

The model assumes constant active power and perfect preservation of completed work at slot boundaries. It omits idle-server power, cooling, networking, storage costs and physical microgrid behaviour.

The next meaningful comparison requires authorised operating data, realistic restart behaviour, agreed service requirements and a competent existing-system baseline.

I want GalaxyLink to retain findings like this openly. An experiment that challenges our first instinct can prevent a much more expensive mistake.

Read the experiment record and the methodology before interpreting any result.

Evidence note: all numerical results come from the GalaxyLink simulator v0.1 validation record and saved demo outputs, dated 1 November 2026.