From the Front Lines of Application Development 8 th : Interview with Kazuya Ishimura, Developer of SMASH

Quantum chemistry calculations in the many-core era

As suggested by its official name (Scalable Molecular Analysis Solver for High  erformance computing systems), SMASH is characterized by the ability to perform electron state and structural optimization calculations as routine work on computers ranging from PC clusters (made up of PCs equipped with scalar CPUs) up to the K computer. In quantum chemistry calculations, approximations based on chemistry and physics knowledge are introduced to the Schrödinger equation ̶ which cannot be solved analytically with the exception of hydrogen and some other atoms ̶ in order to determine the electron distribution of molecules. Specifically, by entering the atomic orbitals (electron distribution of the atom) and coordinates and performing Hartree-Fock calculations (the most basic calculations), it is possible to determine the molecular orbitals (distribution of electrons in the molecule) (Fig.1). The quantity of calculations generally increases in direct proportion to the cube (or, in the case of highly precise calculations, the fifth power or more) of the number of atoms. For this reason, acceleration and parallelization of calculations are essential.
SMASH combines the Pople-Hehre method, which rotates the coordinate axis and reduces the quantity of calculations, with the McMurchie-Davidson method, which uses a recurrence equations to efficiently determine high orbital angular momentum terms. This makes it possible to perform the high-speed atomic orbital 2-electron integral calculations that are performed in almost all quantum chemistry calculations. In addition, redundant coordinates that use bond length, angle and torsion angle information are introduced to reduce the number of geometry optimization cycles, resulting in a more utilitarian program. In actual large-scale parallel calculations on 100,000 CPU cores of the K computer using SMASH, a 50,000 times increase in speed and execution efficiency of 13% were achieved in B3LYP energy calculations for a bilayer graphene molecule (C₁₅₀H₃₀)₂ consisting of 360 atoms (Fig.2). In terms of single-node calculation performance, the Hartree-Fock energy calculation for Taxol ( C₄₇H₅₁NO₁₄) molecule was 40% faster compared with GAMESS.

SMASH represents a ray of light in the large-scale parallel computation of the electron state of enormous molecules, which up to now has been considered to be very difficult. There are high hopes that it will become a basic tool for pursuing efficient development in the many-core era. So what led Dr. Ishimura to decide to tackle the development of a large-scale parallel quantum chemistry calculation program with such enormous potential all by himself?

Background and difficulties faced in the development of SMASH

Dr.Ishimura has been researching quantum chemistry calculations ever since his graduation research days, and he developed the program based on GAMESS. In the process, he has steadily acquired more and more experience in algorithms and program development. His codes of 2-electron integral calculations and the second-order Møller-Plesset perturbation theory (one type of electron correlation calculations) were formally incorporated into GAMESS. But one day, when he tried introducing OpenMP into GAMESS in order to achieve higher parallel performance, he found that major changes were needed, ranging from variable assingments to the basic framework of the program. Ultimately, it became a code that could only perform specific calculations, and naturally it could not be incorporated into the main GAMESS program. Dr. Ishimura became keenly aware of the necessity of the program, which can be performed efficiently in the era of many-core computers. He also realized that there were limits to what could be achieved by development based on existing programs. “So I thought it would ultimately be faster for me to develop a program by myself, from the ground up,” Dr.Ishimura says. And so he began the full-fledged development of a large-scale parallel quantum chemistry calculation program.
When most people hear the phrase “program development,” they undoubtedly picture someone pounding on a keyboard from dawn to dusk. But most of Dr.Ishimura's time is taken up by work with pencil and paper. For him, the process of developing the algorithm through comprehensive consideration of such aspects as developing the form of equations to speed up the program, figuring out the arrangement of multidimensional arrays, designing loops, introducing appropriate approximations and so on is critical. Writing the program itself was not that difficult. On the other hand, even though he was developing SMASH from scratch, he had no idea how it would be evaluated, and that caused him unease and anxiety. The main algorithms had already been published in academic papers, and SMASH was developed to substantiate and make it possible for anyone to freely use those algrorithms. In that sense, developing the program was not an end but a means. So no matter how much time he spent on it as a means to an end, it would be difficult to perceive any output that could be considered an achievement.

New Encounters and Progress

For a long time, Dr. Ishimura worked all by himself to develop SMASH, and now, slowly but surely, his efforts have begun to bear fruit. Since last year when he released the program, he has been contacted by researchers and developers in many fields who want to use SMASH to start performing quantum chemistry calculations or incorporate some of the functions of the SMASH program in the development of their own programs. Unexpectedly, he also received inquiries from people in the field of computer science regarding the details of the program from a computer specialist's standpoint. Because he had decided to distribute under an open source license and had made the program simple to execute, in a mere six months he had succeeded in attracting people in a wide range of fields and producing new seeds for future growth. Dr. Ishimura has great expectations for the power of SMASH users. “With SMASH, you can use just a bit of the program, or you can plunge in and get deeply involved,” he says. “There are a lot of things that need to be done to prepare for a post-K computer world. I'd love for anyone who is interested to get involved in development, and I'm ready and willing to collaborate on a variety of on a variety of levels.”
Meanwhile, SMASH is steadily meeting expectations in terms of quantum chemistry calculation needs. For example, Dr.Ishimura is collaborating with the “Elements Strategy Initiative” of the Ministry of Education, Culture, Sports, Science and Technology, using a quantum chemistry approach to search for alternative metals for automobile exhaust catalysts. In quantum chemistry calculations for metal-oxide supported catalysts, it is thought that a size of at least several nanometers will be needed to ensure that cluster size will have no effect on the properties of the support and catalyst. Dr. Ishimura has already performed electronic state calculations for a Rh catalyst on an AlPO₄ support consisting of 560 atoms (Fig. 3). If more and more of this type of research is accumulated and that combined with computer support enables alternative metal catalysts to be designed, it will greatly reduce the experimental cost involved in catalyst development, and this will make a tremendous contribution to the automobile industry in particular. It is only natural that there are such high hopes for SMASH, which makes possible large-scale parallel quantum chemistry calculations.

Researcher Training in the Post-K Computer Era

Up to now, Dr. Ishimura has developed SMASH all by himself. But in a post-K computer era, he thinks program development will have to be a team effort. To that end, he is actively providing  opportunities to train young people. At a Young Researcher Technical Workshop that Dr.Ishimura organized two years ago, the participants were divided into groups, and each group applied a different fine-tuning to the program based on the profiler data from programs brought in by the participants. The workshop produced visible achievements: reportedly some 20% of the p rticipants achieved a 30% or better performance in computing time as a result of fine-tuning and algorithm review. On the other hand, many of the participants faced the same problems and were stuck on the same points. Accordingly, Dr. Ishimura did not merely seek to improve the skill level of the individual participants but instead provided an environment in which they could pool their knowledge of compiling and fine-tuning. In this way, he focused on working to improve the overall level of the young people in the field.
There are many similarities between developing a software program and training the next generation of young people. Much of the work is done in the background, and a lot of time and effort is required before visible results are produced. It may be because Dr. Ishimura has had the experience of working steadily to develop a software program that he is able to train the next generation using what is, in a sense, a natural process. “It would make me happy if young people would study SMASH just as I once used GAMESS to study the basics of quantum chemistry calculations,” he says. One senses that there is a weight behind these words.
And the acronym SMASH? That comes from the world of badminton, Dr. Ishimura's hobby, which he still practices once a week. Like the shuttlecock in the SMASH logo, we hope SMASH will fly swiftly across the front lines of the post-K computer era.

Kazuya Ishimura   Received a Doctorate of Science from the Graduate University for Advanced Studies and then worked on the development of large-scale parallelization and high-speed algorithm programs for quantum chemistry calculations at Toyota Central R&D Labs., Inc. and Kobe University. The 2-electron integral calculation code that he developed has been incorporated into the GAMESS quantum chemistry calculation program and is still used today as a default routine. Hobby: playing badminton.

◆Interviewer's Postscript Yasushi Shibuta   Through this interview, I realized that there are many points in common between program development, in which enormous amounts of time and effort are required before visible results can be obtained, and the training of the next generation of researchers. As a young university professor, I struggle daily with the question regarding the education of undergraduate and graduate students. It was very inspiring and a great opportunity for me to see the tireless efforts and passion devoted to program development by Dr.Ishimura.