CheckpointIntervalScenario
Reliability scenarios · Implementation: mlsysim/engine/scenario_definitions/reliability.py.
Construct this scenario with keyword inputs. The fields below show the types and defaults declared by this class.
Checkpoints of workload written at write_bandwidth, against failures every mtbf.
write_time states the write directly when no workload is given. Added 2026-10-20.
Notes
Equation: Compose mlsysim.physics.quantities.memory_from_params, mlsysim.solvers.CheckpointIntervalModel, mlsysim.physics.reliability.calc_checkpoint_waste.
Assumptions: Checkpoint writes and the supplied mean failure interval follow the periodic-checkpoint approximation in CheckpointIntervalModel; either a write duration or storage geometry must be provided.
Excludes: Checkpoint contention, changing write rates, and recovery costs outside that solver are absent.
References
Derivation: this scenario binds its input fields to the referenced implementation contracts: mlsysim.physics.quantities.memory_from_params, mlsysim.solvers.CheckpointIntervalModel, mlsysim.physics.reliability.calc_checkpoint_waste. The methods below identify any additional arithmetic.
Examples
>>> from mlsysim import Scenarios
>>> case = Scenarios.Gpt3CheckpointInterval10K
>>> str(case.checkpoint_bytes)
'2450000000000.0 B'Attributes
| Name | Description |
|---|---|
| bytes_per_param | |
| checkpoint_bytes | Persistent checkpoint storage derived from parameter count. Equation: |
| miss_factors | |
| mtbf | |
| result | Evaluate this scenario through mlsysim.solvers.CheckpointIntervalModel.solve. |
| t_write | Checkpoint duration, using an explicit duration when supplied. Equation: otherwise |
| workload | |
| write_bandwidth | |
| write_time |
Methods
| Name | Description |
|---|---|
| miss_penalty | Relative excess waste when the interval differs from the optimum. Equation: |
| waste_at | Expected time fraction lost at a proposed checkpoint interval. Equation: |
miss_penalty
engine.scenario_types.CheckpointIntervalScenario.miss_penalty(factor)Relative excess waste when the interval differs from the optimum. Equation: waste_at(factor * result.interval) / result.waste - 1. Units: dimensionless; the positive factor rescales the modeled optimum.
Extra waste, relative to the optimum, of an interval factor times the optimum (added 2026-10-20); equals (x + 1/x) / 2 - 1, so the minimum is flat.
waste_at
engine.scenario_types.CheckpointIntervalScenario.waste_at(interval)Expected time fraction lost at a proposed checkpoint interval. Equation: mlsysim.physics.reliability.calc_checkpoint_waste(t_write, interval, mtbf). Units: dimensionless; inherits its periodic-checkpoint approximation.
mtbf
Type: Quantity.
Declared default: required.
workload
Type: Optional[Workload].
Declared default: None.
bytes_per_param
Type: Optional[Quantity].
Declared default: None.
write_bandwidth
Type: Optional[Quantity].
Declared default: None.
write_time
Type: Optional[Quantity].
Declared default: None.
miss_factors
Type: Optional[List[float]].
Declared default: None.
checkpoint_bytes
engine.scenario_types.CheckpointIntervalScenario.checkpoint_bytes
Persistent checkpoint storage derived from parameter count. Equation: mlsysim.physics.quantities.memory_from_params(parameters, bytes_per_param). Units: bytes; requires the workload and per-parameter storage convention.
t_write
engine.scenario_types.CheckpointIntervalScenario.t_write
Checkpoint duration, using an explicit duration when supplied. Equation: otherwise checkpoint_bytes / write_bandwidth. Units: seconds; the bandwidth branch omits startup and contention.
result
engine.scenario_types.CheckpointIntervalScenario.result
Evaluate this scenario through mlsysim.solvers.CheckpointIntervalModel.solve. Equation: pass the stated fields to that implementation; its contract defines the calculation and feasibility tests. Units: each returned field follows the referenced result schema; no independent equation is evaluated by this wrapper.