Batch 1: Definition of several Weather & Climate Dwarfs

Gianmarco Mengaldo

REVIEW 3 major objections 4 minor 26 references

Batch 1: Definition of several Weather & Climate Dwarfs

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Six self-contained weather-model kernels are now available as benchmarks.

desk verdict Useful benchmark documentation from the ESCAPE project, but it is not a research paper and the SL-advection dwarf sidesteps the communication cost it is meant to study. read the letter →

arxiv 1908.06089 v1 pith:2IYGEWNK submitted 2019-08-16 cs.DC

Gianmarco Mengaldo This is my paper · ORCID

classification cs.DC

keywords weatherandclimatedwarfsnumericalpredictionhigh-performancecomputingexascalespectraltransformsemi-Lagrangianadvectioncloudmicrophysicsenergyefficiency

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This report defines the first batch of 'weather & climate dwarfs': small, runnable software kernels that isolate the most computationally demanding pieces of operational weather and climate models. Six dwarfs are packaged and documented: the spherical-harmonics spectral transform, a bi-Fourier spectral transform, the IFS cloud microphysics scheme, a generalized conjugate residual elliptic solver, the LAITRI semi-Lagrangian interpolation, and a first version of a semi-Lagrangian advection scheme. Each dwarf comes with installation, testing, and run instructions, and several are extracted directly from the ECMWF Integrated Forecasting System. The aim is to give the ESCAPE project a testbed for measuring time-to-solution, energy-to-solution, and scalability of these kernels on next-generation heterogeneous hardware.

What carries the argument

The central object is the dwarf itself: a self-contained, parallel, readable, verifiable mini-application defined by seven criteria (definition, parallelism, verification, reproducibility, readability, profiling, integration). Each dwarf carries the argument by isolating a specific algorithm: the spherical-harmonics spectral transform pairs FFTs in longitude with Legendre transforms in latitude; the GCR solver uses a preconditioned variational iteration for a three-dimensional elliptic potential-flow problem; the cloud scheme builds a positive-definite matrix equation for cloud/precipitation variables and solves it with an LU-type decomposition; the bi-Fourier transform applies Cooley-Tukey FFTs in both directions with an elliptic spectral truncation and a periodic extension zone; LAITRI performs a 32-point three-dimensional interpolation with precomputed linear, cubic, and vertical weights; and the semi-Lagrangian advection scheme iterates departure-point equations using midpoint or SETTLS extrapolation and then interpolates with 8-, 32-, or 64-point stencils.

What would settle it

Run the spherical-harmonics spectral-transform dwarf at TCo1279 on the same machine as the full IFS model, and compare the measured ratio of MPI transposition time to computation time, together with energy consumed per timestep; if the dwarf's ratio differs from the in-model kernel by more than a few percent, the dwarf's performance conclusions do not transfer to production.

Watch

Extended reading notes

Core claim

The central claim is that a handful of self-contained mini-applications can faithfully represent the dominant computational and communication patterns of operational numerical weather prediction and climate models. The report supports this by identifying six such patterns and providing their scientific formulation, pseudo-algorithms, and working code: two spectral-transform prototypes (one built on the Atlas data structure, one directly extracted from the IFS trans library of cycle CY41R1), a GCR solver for potential flow over a Gaussian hill, the IFS cloud scheme based on an implicit multi-dimensional matrix solver solved level-by-level, the ALADIN-family bi-Fourier transform with elliptic truncation and an extension zone, the LAITRI 32-point quasi-tri-cubic interpolation, and a stand-alone semi-Lagrangian advection scheme with multiple departure-point and interpolation options. Each dwarf is accompanied by regression tests and is intended to be a runnable, verifiable representation of a key functional block, allowing partners to explore algorithmic alternatives and hardware mappings.

Load-bearing premise

The prototypes extracted from production systems reproduce the computational behavior, communication patterns, and energy profile of the same kernels inside the full forecast model, under representative test cases and grid resolutions.

Editorial extensions

If this is right

Each dwarf can be benchmarked on CPUs, GPUs, and optical co-processors to produce comparable time-to-solution and energy-to-solution numbers for a fixed problem.
New algorithmic ideas (fast Legendre transforms, alternative departure-point iterations, quasi-monotone limiters) can be tested in isolation before being folded into the full model.
Because the host versions derive directly from IFS cycle CY41R1 and IFS/ARPEGE/ALADIN cycle 43, optimizations that work in the dwarfs can be merged back into operational code with minimal effort.
The dwarfs cover memory-bandwidth-bound, communication-bound, and compute-bound regimes, so together they provide a compact energy-oriented evaluation of exascale hardware without requiring a full model port.
The set of dwarfs can be extended (radiation, coupling, MPDATA advection) to cover more of the model's total cost envelope in later deliverables.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the dwarf-transfer assumption holds, hardware vendors could pre-screen new architectures on these seven kernels, since the set spans several distinct performance-limiting regimes.
The LAITRI bit-reproducibility requirement shows the project's performance figures are tied to conservative compiler flags; under aggressive optimizations, the time and energy numbers would shift, which would be a separate measurement worth reporting.
A stronger validation of the whole concept would be a blind test: optimize a dwarf for a new architecture, then measure how much of that optimization carries over when the same kernel is exercised inside the full IFS model.
This batch omits coupling dwarfs (atmosphere-ocean, atmosphere-land), so the communication costs of coupled Earth-system models at exascale remain untested; those costs may dominate at extreme scale.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Let me be direct: this is an ESCAPE project deliverable, not a research paper. It describes six mini-apps (“dwarfs”) that capture key computational kernels in weather and climate models, and it ships code with installation instructions and some regression tests. That is genuinely useful engineering documentation for the HPC-for-NWP community.

What is actually new is the packaging: taking established algorithms (spherical harmonics transform, GCR elliptic solver, IFS cloud microphysics, bi-Fourier transform, LAITRI interpolation, semi-Lagrangian advection) and defining them as runnable benchmark units in the Berkeley-dwarf spirit. The algorithm descriptions and pseudo-codes are mostly consistent with the standard literature, and the report is honest about its scope.

The soft spots, in proportion: First, there are no quantitative performance or energy results anywhere in the report. It is a specification, not a benchmark study. That is fine for a deliverable, but it limits what can be evaluated. Second, the cloud microphysics dwarf has no tests yet (“Testing Not supported yet.”). Third, the typesetting has issues: some equations have garbled indices (e.g., Eq. 5 looks like it has a stray bracket; the cloud microphysics matrix in Eq. 24 has subscripts that are hard to parse). These are distracting but not fatal.

The most substantive concern is the semi-Lagrangian advection dwarf. Section 4.6.3.3 states clearly that the current MPI implementation uses a “max wind halo” approach that avoids any data exchange during the trajectory calculation. The production IFS instead uses “insourcing” and “outsourcing” communication-on-demand strategies. So the dwarf does not exercise the very communication overhead the report’s motivation (Section 4.6.1) identifies as the reason to study SL advection at exascale. This means scalability conclusions drawn from this dwarf would not transfer to IFS-class SL schemes. The text acknowledges this by listing the other strategies as future work, so the authors are aware; but as a benchmark, it is a real gap.

Bottom line: this is a useful project report for people working on weather/climate HPC benchmarks or inside ESCAPE. It is not a research contribution and should not be reviewed as one. If someone submits it to a journal as a software paper, I would send it to a referee, but they would need to add validation experiments and address the SL communication gap before I would accept it. For now, treat it as a citable project deliverable, not as a peer-reviewed result.

Referee Report

3 major / 4 minor

Summary. The manuscript is the ESCAPE project deliverable D1.1, defining the first batch of six Weather & Climate dwarfs: the spherical-harmonics spectral transform, the bi-Fourier spectral transform, the GCR elliptic solver, the IFS cloud microphysics scheme, the semi-Lagrangian interpolation routine LAITRI, and a semi-Lagrangian advection scheme. For each dwarf it provides scientific background, governing equations, pseudo-algorithms, I/O interfaces, prototype descriptions, installation and testing instructions, and integration notes. It also gives a formal definition of what constitutes a dwarf and a map of how the dwarfs fit into the dynamical core, physical parametrization, coupling, and time-stepping structure of NWP and climate models. The implementations are available on the project's git platform, and several dwarfs are equipped with ctest regression tests.

Significance. If the code artifacts are faithful to the described algorithms, the report provides a useful modular testbed for exploring algorithm adaptation, benchmarking, and energy-aware optimization for weather and climate models on next-generation hardware. Its strengths are the use of production-derived code (e.g., IFS-based prototypes), the reproducible installation via the `escape` tool, and the generally transparent documentation of test status and algorithmic variants. The report does not present quantitative validation or performance results, so its scientific contribution is primarily as a definitional and enabling document for the ESCAPE workflow.

major comments (3)

[4.3.5.2] Section 4.3.5.2 states 'Testing Not supported yet.' This directly contradicts the VERIFICATION requirement in Section 3 (item 3) and the mandatory ctest policy stated in the Warning boxes of every other dwarf section. As a result, dwarf-P-cloudMicrophysics-IFSScheme is not currently a 'runnable and verifiable mini-application' as defined in Section 3. The authors should either add regression tests for this dwarf or explicitly downgrade it to a non-verified prototype with a clearly visible disclaimer in both the abstract and the dwarf's own section.
[4.6.3.3] Section 4.6.3.3 states that the currently implemented MPI parallelization is a 'max wind halo' approach that requires no data exchange between MPI tasks, while the production IFS 'insourcing' and 'outsourcing' strategies are listed only as future options. This is problematic because Section 4.6.1 motivates the dwarf by the communication overhead that limits SL efficiency at high resolution, and Section 4.6.2 lists 'assess the scalability limits of the SL technique' as a main objective. With only the max-wind-halo variant, the dominant communication-on-demand overhead of production IFS is not exercised, so any scalability or energy conclusions drawn from this dwarf will not transfer to production codes. The report should add an explicit caveat in Sections 4.6.1, 4.6.2, and the conclusions that current results are specific to the max-wind-halo variant until insourcing/outsourcing are implemented.
[4.2 / 4.3 equations] The typesetting of several key equations is corrupted. Eq. (5) is missing a closing bracket (the denominator should be [Pbar^1_N(x_k)]^2). Eq. (24)'s matrix is misaligned with stray opening parentheses, making the linear system difficult to read. In Section 4.2.3.2, the text refers to 'Equation 13' and 'the last equation in Eq. 12' where the intended references appear to be Eq. (8) and Eq. (10), respectively. Since the document's purpose is to define the dwarfs precisely, these equations and references must be corrected before the report can be used as a reliable specification.

minor comments (4)

[4.3.6] The path definition 'inst=dwarf-P-cloudMicrophysics-IFSScheme/sources/' should be '.../install/', since the executable is later invoked from the install directory.
[4.5.4.1] Prototype 1 of the LAITRI dwarf is described as work-in-progress, yet Section 4.5.5.2 states that a testing framework exists and works; please clarify the current status of this prototype.
[4.1.6.2] For prototype2 the report says it should be used as a benchmark but only limited documentation and support are provided; it would help to state explicitly which namelist settings are needed for a standard benchmark run.
[Algorithm 2] In the GCR pseudo-code, the indexing of p^nu and L(p^nu) together with the 'reset' step is ambiguous; a small note linking the pseudo-code to the Fortran subroutine GCR_k would help readers verify the implementation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the dwarfs are documented implementations of established algorithms, with verification against external benchmarks and no fitted-input prediction cycle.

full rationale

This deliverable defines a set of weather and climate dwarf mini-applications and documents how to install, run, and test them. It makes no prediction that is derived from its own fitted parameters: spectral transforms, elliptic solvers, cloud microphysics, bi-Fourier transforms, semi-Lagrangian interpolation, and semi-Lagrangian advection are all implemented from previously published algorithms (e.g., Wedi et al. 2013, Ritchie 1991, Hortal 2002, Temperton 1983) or are directly extracted from operational IFS code cycles. The report explicitly provides external test cases (solid body rotation, Rossby-Haurwitz wave), regression tests, and maximum-error outputs against expected tolerances. The only noted limitations are acknowledged gaps in coverage: the cloud microphysics dwarf has no testing framework yet ('Testing Not supported yet', Section 4.3.5.2), the LAITRI Atlas prototype is 'work-in-progress' (Section 4.5.4.1), and the semi-Lagrangian advection dwarf currently uses a max-wind halo without the production IFS 'insourcing' or 'outsourcing' communication-on-demand strategies, which are explicitly listed as future options (Section 4.6.3.3). These are honest scope limitations concerning transferability of performance conclusions, not circular reasoning: no result is defined in terms of itself, no fitted quantity is relabeled as a prediction, and no load-bearing uniqueness claim rests on a self-citation. The central content is self-contained code documentation and packaging of known algorithms, so the appropriate circularity score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No fitted parameters appear in this report. The central utility of the dwarfs relies on two domain assumptions: representativeness of the extracted prototypes and coverage of key model patterns; both are stated but not demonstrated in the document. The algorithms themselves are standard math drawn from cited prior work.

assumptions (3)

domain assumption The extracted dwarf prototypes faithfully represent the computational and communication behavior of the production NWP models they are taken from.
Section 2.3 and per-dwarf scopes assert the dwarfs mimic key functional blocks, but no validation against full-model profiles is provided.
domain assumption The selected dwarfs cover the dominant computational patterns of weather and climate models.
Section 2.2 describes the dwarf map and notes the selection is partial, with radiation deferred to deliverable D1.4. The current batch is a starting set, so coverage is a stated project assumption.
standard math The cited standard algorithms (FFT, Legendre transforms, GCR, semi-Lagrangian schemes) are correct as described in the referenced literature.
Sections 4.1, 4.2, 4.4, 4.6 invoke prior work (e.g., Temperton 1983, Wedi et al. 2013, Smolarkiewicz and Margolin 1994); the report does not re-derive them.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Batch 1: Definition of several Weather & Climate Dwarfs." pith.science (2026). https://pith.science/paper/2IYGEWNK

@misc{pith2026190806089,
  author       = {Pith},
  title        = {Pith review of: Batch 1: Definition of several Weather & Climate Dwarfs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2IYGEWNK}},
  note         = {Machine review of arXiv:1908.06089}
}

read the original abstract

This document is one of the deliverable reports created for the ESCAPE project. ESCAPE stands for Energy-efficient Scalable Algorithms for Weather Prediction at Exascale. The project develops world-class, extreme-scale computing capabilities for European operational numerical weather prediction and future climate models. This is done by identifying weather & climate dwarfs which are key patterns in terms of computation and communication (in the spirit of the Berkeley dwarfs). These dwarfs are then optimised for different hardware architectures (single and multi-node) and alternative algorithms are explored. Performance portability is addressed through the use of domain specific languages. This deliverable contains the description of the characteristics of the weather & climate dwarfs that form key functional components of prediction models in terms of the science that they encapsulate and in terms of computational cost they impose on the forecast production. The ESCAPE work flow between work packages centres on these dwarfs and hence their selection, their performance assessment, code adaptation and optimization is crucial for the success of the project. At this stage of ESCAPE, a selection of established and new dwarfs has been made, their documentation been compiled and the software been made available on the software exchange platform. The selection of dwarfs will be extended throughout the course of the project (see Deliverable D1.2). The current selection includes the spectral transforms, the cloud microphysics scheme, two and three-dimensional elliptic solvers, a bi-Fourier spectral transform, an interpolation needed for the semi-Lagrangian advection scheme and a first version of the semi-Lagrangian advection scheme itself. This deliverable includes their scientific description and the guidance for installation, execution and testing.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

[1]

Bénard, J

P. Bénard, J. Vivoda, J. Mašek, P. Smolíková, K. Yessad, Ch. Smith, R. Brožková, and J.-F. Geleyn. Dynamical kernel of the Aladin–NH spec- tral limited-area model: Revised formulation and sensitivity experiments. Q. J. R. Meteor. Soc., 136(646):155–169, 2010

work page 2010
[2]

Bermejo and A

R. Bermejo and A. Staniforth. The conversion of semi-Lagrangian advection schemes to quasi-monotone schemes.Mon. Weather Rev., 120:2622–2631, 1992

work page 1992
[3]

D.R. Durran. Numerical Methods for Wave Equations in Geophysical Fluid Dynamics. Springer, 1999

work page 1999
[4]

IFS Documentation, Part IV Physical Processes, Chapter 7 Clouds and Large-scale Precipitation.ECMWF, 2015

ECMWF. IFS Documentation, Part IV Physical Processes, Chapter 7 Clouds and Large-scale Precipitation.ECMWF, 2015

work page 2015
[5]

Eisenstat, H.C

S.C. Eisenstat, H.C. Elman, and M.H. Schultz. Variational iterative methods for nonsymmetric systems of linear equations.SIAM Journal on Numerical Analysis, 20(2):345–357, 1983

work page 1983
[6]

Forbes and A.M

R.M. Forbes and A.M. Tompkins. An improved representation of cloud and precipitation. ECMWF Newsletter No. 129, pages 13–18, 2011

work page 2011
[7]

Forbes, A.M

R.M. Forbes, A.M. Tompkins, and A. Untch. A new prognostic bulk micro- physics scheme for the IFS.ECMWF Tech. Memo. No. 649, 2011

work page 2011
[8]

Haugen and B

J.E. Haugen and B. Machenhauer. A spectral limited-area model formula- tion with time-dependent boundary conditions applied to the shallow-water equations. Mon. Weather Rev., 121(9):2618–2630, 1993

work page 1993

Show all 26 references

[9]

M. Hortal. The development and testing of a new two-time-level semi- Lagrangian scheme (SETTLS) in the ECMWF forecast model.Q.J.R. Meteorol. Soc., 128:1671–1687, 2002

2002
[10]

Malardel and N.P

S. Malardel and N.P. Wedi. How does subgrid-scale parametrization inﬂuence nonlinear spectral energy ﬂuxes in global nwp models?Journal of Geophysical Research: Atmospheres, 2016

2016
[11]

Marras, J.F

S. Marras, J.F. Kelly, M. Moragues, A. Müller, M.A. Kopera, M. Vázquez, F.X. Giraldo, G. Houzeaux, and O. Jorba. A review of element-based Galerkin meth- ods for numerical weather prediction: Finite elements, spectral elements, and discontinuous Galerkin.Archives of Computatio...

2015
[12]

Pudikiewicz, R

J. Pudikiewicz, R. Benoit, and A. Staniforth. Preliminary results from a partial lrtap model based on an existing forecast model.Atmos. Ocean, 23:267–303, 1985

1985
[13]

H. Ritchie. Semi-Lagrangian Advection on a Gaussian grid.Mon. Weather Rev., 115:608–619, 1987

1987
[14]

H. Ritchie. Application of the semi-lagrangian method to a multilevel spectral primitive- equations model.Q. J. R. Meteorol. Soc, 117:91–106, 1991

1991
[15]

Smolarkiewicz and L.G

P.K. Smolarkiewicz and L.G. Margolin. Variational solver for elliptic problems in atmospheric ﬂows.Appl. Math. Comp. Sci, 4(4):527–551, 1994

1994
[16]

Smolarkiewicz and L.G

P.K. Smolarkiewicz and L.G. Margolin. Mpdata: A ﬁnite-diﬀerence solver for geophysical ﬂows.Journal of Computational Physics, 140(2):459–480, 1998

1998
[17]

Smolarkiewicz and L.G

P.K. Smolarkiewicz and L.G. Margolin. Variational methods for elliptic problems in ﬂuid models. InProc. ECMWF Workshop on Developments in numerical methods for very high resolution global models, pages 137–159, 2000

2000
[18]

Smolarkiewicz and J.A

P.K. Smolarkiewicz and J.A. Pudykiewicz. A class of semi-Lagrangian approx- imations for ﬂuids.J. Atmos. Sci., 49:2082–2096, 1992

1992
[19]

Staniforth and J

A. Staniforth and J. Côté. Semi-Lagrangian integration schemes for atmo- spheric models - a review.Mon. Weather Rev., 119:2206–2223, 1991

1991
[20]

Staniforth and J

A. Staniforth and J. Côté. Semi-lagrangian integration schemes for atmospheric models-a review. Monthly weather review, 119(9):2206–2223, 1991

1991
[21]

Temperton

C. Temperton. Self-sorting mixed-radix fast fourier transforms.Journal of computational physics, 52(1):1–23, 1983

1983
[22]

Temperton, M

C. Temperton, M. Hortal, and A. Simmons. A two-time-level semi-Lagrangian global spectral model.Q.J.R. Meteorol. Soc., 127:111–127, 2001

2001
[23]

Thomas, J.P

S.J. Thomas, J.P. Hacker, P.K. Smolarkiewicz, and R.B. Stull. Spectral preconditioners for nonhydrostatic atmospheric models. Monthly weather review, 131(10):2464–2478, 2003

2003
[24]

M. Tiedtke. Representation of clouds in large-scale models.Mon. Wea. Rev., 121:3040–3061, 1993

1993
[25]

N. P. Wedi. Increasing horizontal resolution in numerical weather prediction and climate simulations: illusion or panacea?Philosophical Transactions of the Royal Society A, 372, 2013. 78 ESCAPE - WP1: Weather & Climate Dwarfs - D1.1: Batch no. 1 of Dwarfs

2013
[26]

N.P. Wedi, M. Hamrud, and G. Mozdzynski. A fast spherical harmonics transformforglobalnwpandclimatemodels. Montly Weather Review, 141:3450– 3461, 2013. 79 ESCAPE 2016 Batch 1: Definition of several Weather & Climate Dwarfs Document History Version Author(s) Date Changes 0.1 EC...

2013

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Bénard, J

P. Bénard, J. Vivoda, J. Mašek, P. Smolíková, K. Yessad, Ch. Smith, R. Brožková, and J.-F. Geleyn. Dynamical kernel of the Aladin–NH spec- tral limited-area model: Revised formulation and sensitivity experiments. Q. J. R. Meteor. Soc., 136(646):155–169, 2010

work page 2010

[2] [2]

Bermejo and A

R. Bermejo and A. Staniforth. The conversion of semi-Lagrangian advection schemes to quasi-monotone schemes.Mon. Weather Rev., 120:2622–2631, 1992

work page 1992

[3] [3]

D.R. Durran. Numerical Methods for Wave Equations in Geophysical Fluid Dynamics. Springer, 1999

work page 1999

[4] [4]

IFS Documentation, Part IV Physical Processes, Chapter 7 Clouds and Large-scale Precipitation.ECMWF, 2015

ECMWF. IFS Documentation, Part IV Physical Processes, Chapter 7 Clouds and Large-scale Precipitation.ECMWF, 2015

work page 2015

[5] [5]

Eisenstat, H.C

S.C. Eisenstat, H.C. Elman, and M.H. Schultz. Variational iterative methods for nonsymmetric systems of linear equations.SIAM Journal on Numerical Analysis, 20(2):345–357, 1983

work page 1983

[6] [6]

Forbes and A.M

R.M. Forbes and A.M. Tompkins. An improved representation of cloud and precipitation. ECMWF Newsletter No. 129, pages 13–18, 2011

work page 2011

[7] [7]

Forbes, A.M

R.M. Forbes, A.M. Tompkins, and A. Untch. A new prognostic bulk micro- physics scheme for the IFS.ECMWF Tech. Memo. No. 649, 2011

work page 2011

[8] [8]

Haugen and B

J.E. Haugen and B. Machenhauer. A spectral limited-area model formula- tion with time-dependent boundary conditions applied to the shallow-water equations. Mon. Weather Rev., 121(9):2618–2630, 1993

work page 1993