EpochOverheadScenario
Training scenarios · Implementation: mlsysim/engine/scenario_definitions/training.py.
Construct this scenario with keyword inputs. The fields below show the types and defaults declared by this class.
One training epoch of network on a host CPU and an accelerator: work against fixed overhead.
Every launched operation pays dispatch (the framework’s Python-side cost); on the accelerator it also pays launch. flop_energy and dram_energy_per_byte price arithmetic against data movement.
Notes
Equation: EpochOverheadModel combines batched work, batch-one work, steps per epoch, and per-operation dispatch/launch costs.
Assumptions: The dataset supplies an epoch size and per-operation overhead/energy inputs are held fixed.
Excludes: Measured overlap, dynamic power, and training quality.
References
Derivation: Delegates to mlsysim.engine.solvers.neural.EpochOverheadModel.solve, supplying both the configured-batch and batch-one DenseNetworkModel results. The two profiles let the solver distinguish repeated per-step overhead from work amortized by batching.
Examples
>>> from mlsysim import Scenarios
>>> s = Scenarios.MnistEpochCpuVersusH100
>>> s.network.result.steps_per_epoch > 0
TrueAttributes
| Name | Description |
|---|---|
| accelerator | |
| dispatch | |
| dram_energy_per_byte | |
| flop_energy | |
| host | |
| launch | |
| network | |
| precision | |
| result | Evaluate EpochOverheadModel using batched and single-example network costs and explicit host/launch overheads. |
network
Type: DenseNetworkScenario.
Declared default: required.
host
Type: HardwareNode.
Declared default: required.
accelerator
Type: HardwareNode.
Declared default: required.
precision
Type: str.
Declared default: 'fp32'.
dispatch
Type: Quantity.
Declared default: required.
launch
Type: Quantity.
Declared default: required.
dram_energy_per_byte
Type: Quantity.
Declared default: required.
flop_energy
Type: Quantity.
Declared default: required.
result
engine.scenario_types.EpochOverheadScenario.result
Evaluate EpochOverheadModel using batched and single-example network costs and explicit host/launch overheads.
Equation: Delegates to mlsysim.engine.solvers.neural.EpochOverheadModel.solve with batch and batch-one network profiles.