Performance Tests and Improvements on the rmcdhf and rci Programs of GRASP

Round 1

Reviewer 1 Report

This paper records the results obtained by many users of GRASP seeking to make very large calculations easier and more efficient by using parallelization, MPI and other techniques.  It is impossible for the reviewer to summarize adequately what has been achieved, but it is clear that the authors have made big advances and pointers to making future calculations cheaper.  This gives a clear impression of what can be done with GRASP in its present form, and will help future users to design their computational campaigns.

The English is clear, there are very few typos, and there is very little for the authors to improve, although a conscientious editor can pick out passages that could be better expressed.  I make no suggestions.

Author Response

We feel great thanks for your kind comments.

Reviewer 2 Report

The paper “Performance tests…” by Yan Ting Li and others report resent improvements in the General Relativistic Atomic Structure Package (GRASP). These changes are of the technical nature and improve the efficiency of the package. In particular, more subroutines in the present version are parallelized using the message passing interface (MPI). This allows using modern multicore clusters and (to some extend) improve the scaling with the number of cores. In addition, the authors describe couple of new algorithms, which are implemented in the latest version of the package. The performances of the new and old versions of the package are compared on several examples. The overall speedup significantly depends on the problem and the level of parallelization (the number of MPI processes np). Tests were made for six electron system Mg VII and for four electron Be I and for np from 1 to 48, or more. An average speed up for different basis sets was found to be about one order of magnitude. The scaling with np for the new version is still far from being linear and sometimes is even worse than that for the older version. As a result, the package works efficiently only for the number of processes np about few tens. This means that the package can be effectively used only on clusters of moderate size. The authors argue that the most important next improvement requires more efficient parallelization of the Davidson diagonalization of large sparse matrices.

GRASP is one of the most popular packages for atomic calculations and it is used by many researches worldwide. The paper is addressed mainly to those who work with the package, or considering using it in the future. It may be also interesting to the groups who are developing other atomic packages and account similar problems. For all these researchers the paper will be useful and, I think, deserves publication in the special issue of Atoms devoted to GRASP.

 I have one minor suggestion about the text and more serious one about the figures.

1.       As I said above, the paper is very technical and is difficult to read for those, who never worked with GRASP. However, I think, the style and scope of this paper are suitable for the special issue whose entire readership is interested in the package. I will still suggest to give the unabbreviated name of the package in the abstract.

2.       The authors made rather comprehensive tests comparing old and new versions of the package in different regimes and for different systems. Results of these performance tests are listed in very detailed tables 2-6. In addition, they are illustrated by nine figures. Of course, it is usually easier to look at the plots than into the tables. However, most of these figures, except for Fig. 1 and Fig. 9, are too small. It is necessary to increase the font size and the thickness of the lines on the graphs. I also think that only the speedup plots are essential, while the time plots are excessive and overload the figures.

With these two suggestions I think the paper is ready for publication.

Author Response

We thank the reviewer for these kind words.

Please see the attachment for some responses.

Author Response File: Author Response.pdf

Reviewer 3 Report

The manuscript of Li et al. describes a variety of improvements and optimisations applied to routines within the GRASP atomic structure code. The authors demonstrate that two areas in particular can be improved, namely the multiconfigurational Dirac-Hartree-Fock method and the configuration-interaction routine. The optimisations are achieved using algorithmic improvements as well as the use of parallel programming (message passing interface).

 

The manuscript is well written and provides a detailed account of the relevant numerical methods and algorithms. The results demonstrate the benefits of the work very well. On this basis, I feel that the manuscript is suitable for publication in its current form. 

Author Response

We thank the reviewer for these kind words.