RuntimeGoodputScenario

Serving scenarios · Implementation: mlsysim/engine/scenario_definitions/serving.py.

Construct this scenario with keyword inputs. The fields below show the types and defaults declared by this class.

Runtime goodput of a job on failures.n_devices devices over a held window.

failures is the job’s series failure model (SeriesFailureScenario); the checkpoint is workload at bytes_per_param; recovery is the four-term budget profile (Systems.Reliability.Recovery).

Notes

Equation: runtime result = RuntimeGoodputModel.solve(fleet MTBF, checkpoint bytes, recovery profile, waits, and preemption inputs).

Assumptions: Checkpoint bytes use the supplied per-parameter convention. Device-count sweeps hold all other inputs fixed. Crossover search needs a positive step and a reachable crossing.

Excludes: Searching arbitrary layouts, convergence changes, and a termination guarantee when no crossover exists.

References

Derivation: This configuration forwards its stored inputs to mlsysim.solvers.RuntimeGoodputModel.solve. The solver contract records the equations and their provenance.

Examples

The named scenario supplies illustrative inputs; the output is calculated.

>>> from mlsysim import Scenarios
>>> case = Scenarios.Llama2_70BRuntimeGoodput128
>>> round(float(case.checkpoint_bytes.to('B').magnitude), 3)
980000000000.0

Attributes

Name	Description
bytes_per_param
checkpoint_bytes	Checkpoint bytes from `physics.quantities.memory_from_params` under the declared
failures
input_wait
lost_work
preemption_rate
recovery
result	Modeled result from `mlsysim.solvers.RuntimeGoodputModel.solve` using this
window
workload

Methods

Name	Description
at	Copy the scenario for a new device count while preserving all other job and
crossover	First device count, in positive `step` increments, where failure-side waste
sweep	Map each requested device count to its `RuntimeGoodputModel.solve` result.

at

engine.scenario_types.RuntimeGoodputScenario.at(n_devices)

Copy the scenario for a new device count while preserving all other job and recovery inputs.

Equation: return variant(failures=failures.variant(n_devices=n_devices)).

crossover

engine.scenario_types.RuntimeGoodputScenario.crossover(step)

First device count, in positive step increments, where failure-side waste exceeds input waits.

Equation: Follow the crossover calculation below; branch conditions and rounding are part of its definition.

sweep

engine.scenario_types.RuntimeGoodputScenario.sweep(sizes)

Map each requested device count to its RuntimeGoodputModel.solve result.

Equation: return {n: at(n).result for n in sizes}.