2022/4/8: Original
2022/6/17: Revised to stress importance of not bundling features in a single PR
2022/8/31: Revised to remove duplicate terminology section
2023/4/18: Revised to clarify that modifications to an existing stealth feature is considered a stealth feature.
2023/5/18: Clarify that breaking the addition of a single feature into multiple PRs is welcome
This document describes the policy for bringing in new features and changing existing code in the E3SM model. This
(Implementing this new policy requires a significant amount of work improving: the E3SM diagnostics, CIME test suites and developing documentation, . See collected /wiki/spaces/CNCL/pages/3315138566)
Table of Contents | ||||
---|---|---|---|---|
|
...
BFB (bit-for-bit)
New code produces results which are bit-for-bit identical with old code
Does not contain stealth features
github PR label: BFB
Roundoff
New code would match old code in exact arithmetic, but will diverge exponentially fast when using finite precision
Climate of new and old code will converge as the averaging time goes to infinity
github PR label: non-BFB
Climate changing
New feature is brought into the model turned on
Other model configuration change (parameters, resolution, forcing data)
github PR label: CC
Stealth Feature
BFB code change that includes a new feature, turned off by default, or modifications to an existing stealth feature.
github PR label: Stealth
labels Stealth
Note: multiple PR labels are possible. A Stealth feature is likely also BFB when the feature is off. CC, BFB and non-BFB are mutually exclusive.
Reference Solutions
The E3SM project will maintain reference solutions reflecting the current state of the ‘main’ branch
Coupled model: B-case runs (~100 years) for WC, cryosphere and BGC configurations
Component models: F, I and G cases, typically much shorter (e.g. 5 year F case with cyclic year 2010 forcing)
Updated periodically when climate changes are integrated, or monthly ( to check for unintended changes)
New features will be evaluated with respect to these reference solutions, with documentation on the import metrics
Note: E3SM Reference solutions and related documentation not yet ready!
Section 1: New Feature Request Process and Documentation Requirements
For new features or other major code changes, prepare a New Feature Overview Document.
(To determine when a code change becomes a major code change, consult with the component lead.)
E3SM-SFA funded work: feature should appear in that group’s roadmap before starting any work on a new feature including documentation prep. (E3SM pre-processing code (compsets, cime) and infrastructure are also covered by these procedures - infrastructure group is the “component group” for this work)
Externally funded work: developers are encouraged to submit this document early in their development process. This will assist with later E3SM acceptance.
Developer: Before writing code: Complete Proposer completes the New Feature Overview Document. Include, as appropriate:
High-level description of code changes and/or new code, an overview of the design, infrastructure changes
Expected improvements and how these will be demonstrated
Describe needed updates to E3SM documentation
Expected impacts on computational performance and mass/energy budgets.
If relevant: describe papers that will be published
E3SM Staff: Before writing code: Document Review:
Reviewed byReviewer within E3SM (component lead or their delegate
. Reviewer) should:
Review document for completeness
Determine if there is sufficient benefit to E3SM to justify the E3SM integration and future maintenance costs.
For features not needed by the E3SM SFA, but which may be needed to support other DOE BER missions, the component lead should consult with E3SM leadership.
Determine if review by performance and infrastructure groups
ifare needed and ensure they are done.
discourage and/or flag use of advanced language features or unnecessary complexity that may not be supported well by compilers or may cause performance degradation
Create
onea confluence page
per document to allowfor discussion during the review process
Confluence space for all new feature overview documents and their approval status.
Developer: After new feature is approved, begin work in an E3SM fork, following Development Getting Started Guide
Features which are not approved but are needed for other reasons can be maintained by the developer on their E3SM fork.
Large new features should be developed through a series of PRs rather than a single massive PR (which would be hard to thoroughly test and scrutinize). One New Feature Overview Document can be used for all PRs associated with that feature.
Developer: When work is completed:
Revise overview document to show improvements as originally proposed in overview document, and document simulation results from new feature PR guidelines. Include links to as yet to be constructed results archiving system (once it is ready)
Submit PR PRs with links to documentation, follow new feature PR process.
...
Provide evidence that the changes are roundoff:.
e.g. subcomponent tests which have analytic solutions and error metrics. errors should change only at the roundoff level
e.g. subcomponent tests which do not have exponential growth of roundoff error: in standalone dycore tests roundoff level changes often result in only roundoff level differences in the output.
If roundoff-inducing changes can be isolated to a small amount of code, it might be possible to verify by inspection (e.g. changing order of operations)
Run nbfb tests. If any of them fail, additional scrutiny is warranted.
Strong evidence of roundoff level changes is sufficient in order to not needlessly delay our development process. In the rare case that the change is in fact not roundoff, this will be caught by our monthly reference B case update.
...
Stealth Features. The developer needs to run several tests (with the feature turned on if its a stealth feature) and note in the PR that they have:
...
New and stealth features: dont don't submit PR until Section1 Section 1 approval of overview document
Submit the PR to the E3SM github site following Development Getting Started Guide . Links to the material outlined in Section 2 above should be in the first comment after the PR description.
For code with potential performance impacts: Performance group lead or delegate reviews supplied PACE data.
It is important not to bundle multiple features or fixes into a single PR.
Climate Changing PRs (new features and bug fixes)
Simulation group lead approves based on climate changes
Simulation group lead (or delegate) will be assigned as a reviewer as well as appropriate component lead and then verifies:
source code review was completed
component lead also approves
PR includes test (or is covered by existing tests)
climate changing evaluation process (from Section 2) completed and results archived
Updates main branch reference B case as needed.
PR contains a single feature ( to ensure multiple features are tested independently)
Stealth features
Component lead (or delegate) is assigned as a reviewer and verifies:
source code review
PR includes appropriate test
stealth-feature evaluation process (from Section 2) completed and results archived
Expanded BFB tests were run with feature turned on
PR contains a single feature ( to ensure multiple features are tested independently)
Roundoff level PRs
Component lead (or delegate) verifies:
source code review was completed
evidence to support roundoff level tests is sufficient, if not, follow component new feature climate evaluation procedures
no stealth features
BFB
component lead (or delegate) be assigned as a reviewer verifies:
source code review was completed
no stealth features
...
There is a suite for each major component model and you should. ( atm
, ocn
, ice
, lnd
, rof
, and fully-coupled wcycl
). They cover the standand-resolution grids in both optimization and debug configurations.
See tests.py for details on the included tests.