LISFLOOD-FP 8.1: New GPU accelerated solvers for faster fluvial/pluvial flood simulations
- 1Department of Civil and Structural Engineering, The University of Sheffield, Western Bank, Sheffield, UK
- 2Risk Management Solutions Ltd., London, UK
- 3School of Geographical Sciences, University of Bristol, Bristol, UK
- 1Department of Civil and Structural Engineering, The University of Sheffield, Western Bank, Sheffield, UK
- 2Risk Management Solutions Ltd., London, UK
- 3School of Geographical Sciences, University of Bristol, Bristol, UK
Abstract. The local inertial two-dimensional (2D) flow model on LISFLOOD-FP, so-called ACC uniform grid solver, has been widely used to support fast, computationally efficient fluvial/pluvial flood simulations. This paper describes new releases, on LISFLOOD-FP 8.1, for parallelised flood simulations on the graphical processing units (GPU) to boost efficiency of the existing parallelised ACC solver on the central processing units (CPU) and enhance it further by enabling a new non-uniform grid version. The non-uniform solver generates its grid using the multiresolution analysis (MRA) of the multiwavelets (MWs) to a Galerkin projection of the digital elevation model (DEM). This sensibly coarsens the resolutions where the local topographic details are below an error threshold ε and allows to properly adapt classes of land use. Both the grid generator and the adapted ACC solver on the non-uniform grid are implemented in a GPU new codebase, using the indexing of Z-order curves alongside a parallel tree traversal approach. The efficiency performance of the GPU parallelised uniform and non-uniform grid solvers are assessed for five case studies, where the accuracy of the latter is explored for ε = 10-4 and 10-3 in terms of how close it can reproduce the prediction of the former.
Mohammad Kazem Sharifian et al.
Status: final response (author comments only)
-
RC1: 'Comment on gmd-2022-259', Anonymous Referee #1, 21 Nov 2022
The manuscript titled "LISFLOOD-FP 8.1: New GPU accelerated solvers for faster fluvial/pluvial flood simulations" deals with the upgrade if the well-known LISFLOOD hydrodynamic simulator, using parallel programming and specifically the GPU capabilities in order to speed up the simulations. Except of the parallelization, the authors deomnstrate the use of a smart grid coarsening way, which also speeds up the simulations but with an accuracy sacrify. The paper is well written and well structured and characterized by novelties. I would suggest to be published after some minor technical corrections:
1) It is not consistent to compare all the numerical results (uniform, non-uniform 10^-3, non-uniform 10^-4) against the observed data. Since the non-uniform is an simplification of the uniform detailed grid, the latter should be the base of comparison and the observed values should be given as a supplementary material, not substantial for the core of the paper. The situation in which the non-uniform grid performs better than the uniform grid is rather a coincidence. I assume that the non-uniform grids introduce a kind of artificial diffusion, while similar results could be derived by the uniform grid with bigger values of Manning coefficients.
2) In L335-340 the authors state that a possible cause of the discrepancy between the modelled and the observed hydrograph is the low Reynolds numbers of the flow. However flow ranges between 20 and 100 m^3/s. With these values is impossible to have low Reynolds numbers in the channel. The authors probably mean the rainfall-driven overland flow in the catchment and not in the hydrographic network.Â
3) I really appreciate that the authors are not charatcerized by arrogancy and they give very rational conclusions avoiding global suggestions. However since the paper is mainly demonstrates new tools it might be better to give a more clear practice guidance for the modeller and how to handle every DTM resolution. A table with these suggestions might be good alternative which also highlights the main findings of the work.
Â
-
RC2: 'Comment on gmd-2022-259', Ilhan Özgen-Xian, 29 Nov 2022
The authors present a GPU-accelerated shallow flow solver that is being incorporated into the well-known LISFLOOD framework. Â This effort to improve on existing and established software used by many practicioners is a timely and relevant endeavour. Â The improvements presented in the manuscript are (i) GPU-acceleration with specific focus on non-uniform Cartesian grids and (ii) wavelet-based mesh adaptation to generate these grids. Â The manuscript is well written and easy to follow. Â The selected test cases are meaningful and support the authors' claims.
Therefore, I suggest accepting the manuscript after minor revisions.
1. Governing equations
1.1 The authors omit showing the governing equations.  At least for me, it made the introduction of this paper difficult to follow, specifically the discussion of the ACC solver and its differences to the fully dynamic shallow flow solver (page 1, lines 35—45).  I would suggest showing the SWE explicitly, naming the acceleration term and momentum terms in these equations, and then showing which terms drop out in the ACC solver.
1.2 The acronym ACC is being used without explanation. Â Please provide the full name of this solver the first time you use the acronym.
2. Morton codes
2.1 Out of curiosity, what upper bound does the use of Morton codes give you for the allowable number of cells? Â I am asking because, if I understood correctly, if you combine the binary representation of two integers into one, you can only use half the length that a computer can store for each representation. Â So on a 64 bit machine, this would mean that the maximum number of elements that can be used is what can be represented with 32 bits. Â Is this correct?
3. Lower Triangle catchment
3.1 Source of data should be completed with the source of the raw data:
  Wainwright, H. and Kenneth, W. (2017). LiDAR collection in August 2015 over the East River Watershed, Colorado, USA. https://doi.org/10.21952/WTR/14125424. Fluvial vs. pluvial test cases
Perhaps some of the answers to the comments below could be placed in the Conclusions and recommendations section.
4.1 From the test cases and my own experience, mesh coarsening seems to "work better" for fluvial runoff, probably because pluvial runoff yields very small water depths that elevate the influence of the topography. Â This is to some extent supported by the authors' results. Â Can the authors comment?
4.2 The hydrograph of the Upper Lee catchment shows that coarser grids damp short time-scale events. Â This has been my experience with multiresolution meshes as well. Â Are there mitigations the authors suggest that could lead to more accurately capturing these short time-scale events?
4.3 The pluvial flooding in the Glasgow urban area is very well captured, compared to Lower Triangle and Lee catchments. Â Is this due to the regularity of the urban area?
Â
Mohammad Kazem Sharifian et al.
Mohammad Kazem Sharifian et al.
Viewed
HTML | XML | Total | BibTeX | EndNote | |
---|---|---|---|---|---|
465 | 140 | 12 | 617 | 5 | 5 |
- HTML: 465
- PDF: 140
- XML: 12
- Total: 617
- BibTeX: 5
- EndNote: 5
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1