Analyse a Run's Observed Compute Cost

Reads the per-task .start/.end/.host timings a completed run left in its fit/hc shard Parquets (the cost-analysis instrumentation), and attributes the observed compute to the scenario's ci_method x nboot axes - the measured counterpart to ssd_estimate_cost()'s prediction. It is strictly read-only: it reads result Parquets (and, optionally, a targets meta store), and never runs the pipeline, fits a distribution, draws random numbers, or writes a file. The observed total is serial-equivalent compute (the sum of per-task durations), distinct from elapsed wall time under parallel workers.

Usage

ssd_analyse_cost(scenario, root = scenario_results_dir(scenario), store = NULL)

Arguments

scenario

An ssdsims_scenario from ssd_define_scenario().

root

The run's results root (the fit/hc shard trees live under it). Defaults to scenario_results_dir() for scenario.

store

Optional path to the run's targets data store, to add the per-shard envelope overhead. NULL (the default) reads the shards only.

Value

An ssdsims_cost_analysis: a list with total and longest (both difftime), a breakdown tibble grouped by ci_method x nboot, fit_seconds (the measured fit addend), the hosts seen, a measured flag, the per-shard envelope (when a store is given), and provenance.

Details

Given a targets store, it additionally reads each shard target's wall seconds from targets::tar_meta() and reports the per-shard envelope overhead (target seconds - sum(task durations): parent read, Parquet write, and dispatch), the number that informs partition_by tuning. The combined summary and upload_<step> targets are excluded; errored/unbuilt (NA-seconds) targets are dropped from totals; unmatched stored targets are reported, never silently dropped.

See also

ssd_estimate_cost(), ssd_compare_cost(), ssd_calibrate_cost_from_run().