Performance#
Autosubmit includes built-in support for CPMIP (Computational Performance Model Intercomparison Project) metrics, a set of metrics that can be used for the study of computational performance of climate (and Earth system) models. These metrics provide a quick and comparable view of how efficiently a simulation runs on a given platform, making them a useful first indicator when investigating for inefficiencies — undersized partitions, oversubscribed nodes, I/O bottlenecks, or regressions after a code change.
CPMIP metrics target the simulation job of an experiment — the job that
advances model time chunk by chunk (commonly named SIM). SYPD and CHSY are
defined in terms of simulated years, so they rely on the chunk calendar
(CHUNK, CHUNKSIZE, CHUNKSIZEUNIT) that only chunk-based simulation
jobs carry; CORE_HOURS can still be computed on any job.
CPMIP metrics are not computed automatically: today they are evaluated only as part of the notification workflow. A job’s CPMIP metrics are computed only when both of the following conditions are met:
MAIL.NOTIFICATIONS: Trueis set in the experiment configuration, andthe job declares a
CPMIP_THRESHOLDSblock injobs_<EXPID>.yml.
Without that, the metric values are not recorded anywhere — there is no silent computation in the background. See How to configure CPMIP threshold notifications for how to enable both.
Available CPMIP metrics#
Metric |
Definition |
Formula |
|---|---|---|
SYPD |
Simulated Years Per Day — how much model time the job advances per wall-clock day. Higher is better. |
|
CHSY |
Core-Hours per Simulated Year — total CPU time consumed per year of simulation. Lower is better. |
|
CORE_HOURS |
Total core-hours billed by the run. Useful as a budget signal. |
|
Inputs come from the job stat file (start_time, end_time) and the
experiment configuration (PROCESSORS plus the chunk calendar). If a required
input is missing, the metric is silently skipped — for example, CORE_HOURS
and CHSY are skipped when PROCESSORS is not set on the job.