Why checkpointing changes energy-aware scheduling

A scheduling decision can look efficient until we ask what happens to the work already in progress.

For our first research scaffold, checkpointing means a deliberately simple assumption: completed useful work is preserved at the boundary between time slots. An interrupted job can later continue from that saved progress.

Preserving progress does not make interruption free in the model.

The rule we chose

After an interruption, a job incurs a fixed restart delay. During those restart slots, it consumes its full modelled power but completes no useful work.

If it is interrupted again while restarting, the full delay begins again. A policy that changes its mind too often can therefore spend a significant part of the replay preparing jobs to work.

This is one reason the initial experiment produced a poor outcome for the simple power-packing heuristic. Its decisions created repeated interruptions and substantial modelled restart effort.

What this assumption makes visible

The model separates energy spent from useful progress.

That distinction matters when interpreting a result. A machine appearing active is not sufficient evidence that it is delivering valuable work. Likewise, moving a job to a period with more available capacity may be unhelpful if the change introduces enough delay to miss its deadline.

These observations follow from the chosen model. Their importance in an actual environment must be measured.

What the assumption leaves out

Perfect preservation at every slot boundary is optimistic. Our scaffold does not model the cost of writing a checkpoint, its size, the storage it requires or the consequences of an unusable saved state.

It also assumes constant active power, fixed work units and a fixed restart duration. Those choices make the example understandable. They limit what its outputs can establish.

A future real-data study should ask how often progress can be saved, what that process consumes, how recovery behaves and which jobs should never be interrupted.

The answers may vary between workloads. The study should retain those differences rather than assigning one convenient recovery cost to everything.

The comparison we should require

Every scheduling policy should face the same recovery assumptions and service requirements. If an improved result depends on treating its interruptions more generously, the comparison is unfair.

We should also examine unfinished and late work, not simply compare energy consumed during the replay.

My aim is to make such assumptions visible early. A useful research tool helps us identify the measurements we need before making a promise about performance.

Our methodology explains the broader approach. The library connects this question to constrained computing and infrastructure reliability.

Evidence note: the model description is based on the GalaxyLink simulator v0.1 README and validation record, dated 1 November 2026.