Accelerated Parallel Numerical Simulation of Large-Scale Nuclear Reactor Thermal Hydraulic Models by Renumbering Methods

Round 1

Reviewer 1 Report

The manuscript presents a renumbering method that aims to accelearte the parallel CFD simulations. The topic is intesesting and the methodology is solid. The benefit of this renumbering method is presented based on a test on an in-house general-purpose CFD software called YHACT. It would be even convincing to the scientific community if the authors test the presented renumbering method on some other general-purpose CFD tools. Those CFD tools can either be certain commercial software like Fluent or Star-CCM+ or some open-source CFD codes like OpenFOAM. 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

In their manuscript, the authors describe their attempts to increase the efficiency of parallel CFD simulations with grid renumbering and process partitioning in their YHACT open source software. They describe in some detail the challenges with process communication in the parallel computation and their implementation, and they demonstrate their method with a demonstration for nuclear fusion reactors.

 

The writing is generally pretty clear, but I have some specific suggestions for improvement listed below. My main concern is with the novelty of the results. The reference [27] (by the same authors and published in the same journal) already describes the grid renumbering approach in YHACT and demonstrates the results on the same thero-hydraulic problem. As far as I can tell, the main differences in the current work are i) the specific RCM renumbering algorithm in the current manuscript (which shows only a moderate improvement over the greedy algorithm) and ii) finer meshes in the numerical demonstration. I’m not sure that this is sufficiently important to justify a second publication in the same year, but I will leave that to the editors to consider. The authors should emphasize any additional contributions in the introduction.

 

Specific comments:

 

1)    General comment: punctation at the end of sentences is frequently wrong. Examples include a) no spaces separating periods from the following sentence, b) incomplete introductory clauses that should end in comma rather than period, c) capitalization errors following periods. Proofreading is necessary.

2)    Several acronyms are not defined in their introduction. At the very least, please describe what YHACT and RCM stand for when they are first mentioned.

3)    Page 2: isn’t the idea of heuristic algorithms that they will take less time to train or iterate? Please describe what specifically qualifies refs [21] and [22] as “heuristic” in comparison to [16-20].

4)    Section 2.1 – there is excessively long and repetitive discussion about the benefits of the YHACT implementation. Figure 1(a) is entirely redundant – Figure 1(b) already emphasizes the new contribution and is self-contained.

5)    Section 2.2.1 – the section describes sub-cells as non-overlapping, but Figure 2 depicts overlapping cells for the four central nodes. Also, what is a meant by a “piece” of grid data?

6)    Section 2.2.2 – Figure 3, the second “2” in the row corresponding to “0” should be grey, as other “received” communications are.

7)    Section 2.2.3 describes a “further” partitioning in addition to the “sub-grids” defined in 2.2.1. The new convention of using bold words to define “regions” and “zones” is confusing, since it was not used for “sub-cells” above. Also, it is unclear if there is any relation between “sub-cells” and “regions,” since regions may contain multiple sub-cells according to 2.2.4. Do I understand correctly that parMetis from Section 3 is used to define the “sub-cells” rather than the “regions”, while the “regions” are selected based on the computer architecture? This should be clearer in the manuscript.

8)    Section 2.2.3, what is meant by coprocessor? Does that refer to a node with multiple cores, or a FPU or GPU? Since the text only refers to OpenMP, I assumed there are only compute cores, and no “traditional” coprocessors. Also, for compute nodes, my understanding is that BLAS vectorization is often more beneficial than OpenMP.

9)    Figure 4 and 5 have some dissonance. In figure 4, the each region contains Zones 0,1,2, while Figure 5 seems to suggest the zones in Proc j start at M+1 rather than 0.

10) Section 3.2 – I don’t understand “the number of nonzero elements in each row is only as many as the number of elements associated with that row.” Isn’t the number of non-zero elements exactly equal to the number of elements associated with the row?

11) Section 3.2 – aren’t Gaussian elimination and LU not considered iterative methods? Isn’t the traditional “direct” method exactly the LU method?

12) Section 3.2.2 – the “sparsity” is usually defined as the proportion of non-zero elements. Isn’t the “sparsity” independent of the grid-renumbering method? I think the “bandedness”, the “dispersion”, or even the “sparsity structure” can be said to differ after grid renumbering, but the sparsity itself does not.

13) Section 4.1 – “the solid models of fuel rod and control rod are neglected.” Does this mean they are not detailed in the manuscript, or they are not modeled at all in the numerics? What boundary conditions are used at the solid interfaces for the energy transport?

14) The MDMP “discriminant” is defined in Section 3.2.2. Is this the same as the MDMP “metric” and MDMP “index” in section 4.2? Also, what do the “serial means” refer to in 4.2?

15) Section 4.2 – the original matrix before renumbering is said to be drawn from the ICEM library. Is this different from the Parmetis routines described in 3.1?

16) The “robustness” tests for the numerical algorithms described in 4.3.1 may be inappropriate for chaotic dynamics. The exact value of the fields in two numerical simulations can diverge because of chaotic dynamics simply because of the sensitivity to initial conditions and floating point precision, even if both simulations have good error bounds and are correct.  In any case, it seems obvious that the grid renumbering procedure will not invalidate the numerics if it is implemented correctly.

17) Section 4.3.2 – the table caption and the text in the section are redundant. Include one or the other.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Review. The manuscript discusses the improvement of computational performance of an in-house CFD software YHACT by applying a mesh renumbering technique. The optimal renumbering technique is chosen from a set of known methods. It is demonstrated that the added renumbering technique does not alter the physical results of the model and that it accelerates the calculation across a wide range of parallelization levels.

4.3.1. Proof of Correctness. Perhaps a more appropriate comparison technique can be used to assess the (in)difference between the “old” and “RCM” cases instead of visual comparison of physical quantity distributions? Relative error when subtracting the field values?

4.3.2. Parallel testing of large-scale numerical simulations. Table X should be assigned a number.

Conclusions. The statement “Thus the goodness of the strong scalability of parallelism was tested and proved.” should ideally be proceeded by a clear supporting statement from which the “goodness of the strong scalability” can be concluded.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf