WAVETRISK-1.0: an adaptive wavelet hydrostatic dynamical core

Kevlahan, Nicholas K.-R.; Dubos, Thomas

doi:https://doi.org/10.5194/gmd-12-4901-2019

Articles | Volume 12, issue 11

https://doi.org/10.5194/gmd-12-4901-2019

© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.

https://doi.org/10.5194/gmd-12-4901-2019

© Author(s) 2019. This work is distributed under
the Creative Commons Attribution 4.0 License.

Articles | Volume 12, issue 11

Model description paper

|

27 Nov 2019

Model description paper |

| 27 Nov 2019

WAVETRISK-1.0: an adaptive wavelet hydrostatic dynamical core

Nicholas K.-R. Kevlahan and Thomas Dubos

Download

Final revised paper (published on 27 Nov 2019)
Preprint (discussion started on 16 May 2019)

Interactive discussion

Status: closed

AC: Author comment | RC: Referee comment | SC: Short comment | EC: Editor comment

- Printer-friendly version

- Supplement

SC1: 'Information on speed performance is missing', Antoon van Hooft, 30 May 2019
- AC1: 'Response to comment on performance and other issues', Kevlahan Nicholas, 12 Jun 2019
RC1: 'Review of WAVETRISK-1.0: an adaptive wavelet hydrostatic dynamical core', Anonymous Referee #1, 02 Jul 2019
- AC2: 'Response to Referee 1', Kevlahan Nicholas, 24 Sep 2019
RC2: 'Review', Anonymous Referee #2, 28 Aug 2019
- AC3: 'Response to Referee 2', Kevlahan Nicholas, 24 Sep 2019

Peer-review completion

AR: Author's response | RR: Referee report | ED: Editor decision

AR by Nicholas Kevlahan on behalf of the Authors (24 Sep 2019) Author's response

ED: Referee Nomination & Report Request started (26 Sep 2019) by Paul Ullrich

RR by Anonymous Referee #1 (09 Oct 2019)

Suggestions for revision or reasons for rejection

I would first like to thank the authors for addressing my
comments. The reporting of absolute performance which was added in the
revisions is important since improving the accuracy/cost ratio is the
primary aim of adaptive methods.

This reporting should still be improved however. There seems to be
several inconsistencies which must be clarified before the paper can
be published. See in particular point 6 below.

1) Since the computational cost can be assumed to be proportional to
the number of levels, it should be reported as
cpu/(columns*level/core) rather than cpu/(columns/core). This would
allow direct comparison between benchmarks using a different number of
levels.

2) Table 2 is useful but should also include the cost i.e. an
additional column with something like

Rossby wave 0.86*160/1.94e5 = 0.7 ms /27 0.026 ms
Baroclinic instability 0.40*40/1.64e5 = 0.1 ms /27 0.0037 ms
Baroclinic instability 1.13*40/4.07e5 = 0.1 ms /27 0.0037 ms
Held-Suarez 1 degree 0.27*40/1.11e5 = 0.1 ms /19 0.005 ms
Held–Suarez (1/4 ) 0.29*320/4.54e5 = 0.2 ms /19 0.01 ms

3) The data in the Table above does not seem consistent with the data
reported in Figure 4, right, with values around 1 ms, rather than 0.1
ms here. This must be checked/clarified. Similarly, why is the Rossby
wave case so "expensive"?

4) p. 25 "the high resolution case is run on 640 cores", yet Table 2 says
320 cores. Which is it?

5) To avoid any ambiguity "milliseconds" should be spelled out somewhere.

6) The most favourable case of 0.0037 ms (per timestep per degree of
freedom) does not compare well at all with other implementations of
comparable numerical schemes. For example for the hydrostatic,
sigma-coordinate, ROMS code, this cost was reported to be 2e-5 ms
(almost 200 times faster!) and this back in 2005/2007... (I didn't
look for more recent numbers) see e.g.

Zuo, Yue, Xingfu Wu, and Valerie Taylor. "Performance Analysis and
Optimization of the Regional Ocean Model System on TeraGrid."
TeraGrid'07 Conference.

Wang, Ping, et al. "Parallel computation of the regional ocean
modeling system." The International Journal of High Performance
Computing Applications 19.4 (2005): 375-385.

Even though ROMS has been highly optimised, one would expect even
naive implementations of explicit shallow-water solvers to have
absolute performance of order at least 1e-4 ms (per timestep per
degree of freedom, /dtdf).

I sincerely hope that this is due to a trivial error in the numbers
above. If this is not, then this must be discussed in detail
(i.e. possible explanations and solutions should be proposed) and the
conclusions on the usefulness/applicability of the
method/implementation should be toned down (a lot). Note that the
numbers given above for ROMS are indicative only (more recent numbers
would most probably be even less favourable) and the authors should
look for performance data for other (atmospheric) solvers.

A few more minor comments follow:

p.23 "Nevertheless, figure 4 (left) shows reasonably linear speed up
to at least 640 cores."

"The more sensitive measure of strong scaling in figure 4 (right) is
further from a constant (with a maximum variation of 3 times),
although there is not a definite trend with increasing numbers of
cores."

The scaling for the adaptive case in figure 4, left or right, is
definitely not linear. Figure 4-right is not "a more sensitive
measure" than figure 4-left (it is exactly the same data), it just
leaves less room for wishful interpretation. It is obvious that the
two sets of points on figure 4-left have different slopes. The
non-adaptive case indeed shows linear scaling (from 8 cores up), as
also confirmed more clearly on figure 4-right, while the adaptive
case shows a roughly power-law scaling with an exponent clearly lower
than 1 (maybe even 1/2? it would be good to indicate this), as more
clearly shown on figure 4-right.

p.28 "We demonstrate excellent strong parallel scaling up to at least
2560 cores in the perfectly balanced case and up to at least 640 cores
in an unbalanced case." The second part of this sentence needs to be
modified to reflect the previous comment.

Hide

ED: Publish subject to minor revisions (review by editor) (09 Oct 2019) by Paul Ullrich

AR by Nicholas Kevlahan on behalf of the Authors (16 Oct 2019) Manuscript

ED: Publish as is (25 Oct 2019) by Paul Ullrich

AR by Nicholas Kevlahan on behalf of the Authors (28 Oct 2019) Manuscript

Short summary

WAVETRISK-1.0 is a new adaptive dynamical core for global climate modelling. It uses multiscale adaptive wavelet methods to adjust the grid resolution of the model at each time to guarantee error and make optimal use of computational resources. This technique has the potential to make climate simulations more accurate and allow much higher local resolutions. This "zoom" capability could also be used to focus on significant phenomena (such as hurricanes) or particular regions of the Earth.