Multi-physics Extension of OpenFMO Framework

Multi-physics simulations are widely used even in complex scientific studies. Such calculations are often constructed by combining multiple theories including different degrees of approximations and different scales of description. Since reality and accuracy are required increasingly, these simulations have become larger and more complicated year by year. Grids, distributed computer resources over wide-area networks, are expected to execute such complicated scientific applications, and have been installed all over the world in order to demonstrate large-scale heterogeneous simulations with the help of middlewares [1,2]. On the other hand, the next generation supercomputer with a petascale performance is already planned in several countries [3,4]. Thus, the development of high-performance computing environments is fast and transient. As a scientist, it is important to watch the trend of those computer resources.

In the present contribution, the multi-physics calculations by Fragment Molecular Orbital (FMO) method [5] are constructed on the distributed computing environments. OpenFMO framework toward “peta-scale” computing [6,7] is extended to the multi-physics simulations. It is also discussed what architecture and development policy should be chosen in the fast-moving world of computing.

Before entering the main subject, we briefly review the grid-enabled FMO implementations developed in the NAREGI project [1]. These are based on the famous MO package, GAMESS [8].

Although it is usually considered as an approximation to ab initio molecular orbital (MO) calculations, the FMO algorithm is a multi-layered problem (see Fig. 1(b)) including the MO calculations for each fragment and the electrostatic (ES) interaction between fragments. In the MO-layer, the quantum mechanical interactions of all the atoms and electrons within a fragment are included to obtain a fragment energy. On the other hand, only the classical ES interaction is considered when we go over the boundary of fragments. Since the MO-layer calculations can be executed independently, we can break the program into loosely-coupled components corresponding to a large-scale parallel execution in the distributed computing environments (Fig. 1(c)). The grid-enabled version called “Loosely-coupled FMO” was developed as a part of NAREGI [1]. The total control flow is constructed by the use of the NAREGI Workflow tool. In Fig. 2 (a), the total electron density of the whole molecule [9] of a Gramicidin-A is shown as an equi-density surface, and the electron density for one of the fragments in a fatty-acid albumin is shown in Fig. 2(b).

As an example of the multi-physics simulations, a coupled simulation of FMO and 3D-RISM is presented, where FMO calculations are coupled to statistical mechanics calculations for molecular liquids by Reference Interaction Site Model (RISM) [10]. In order to obtain properties of bio-molecules, drugs, enzymes, etc., it is necessary to perform calculations under the influence of a solvent since these molecules usually work in aqueous solution. However, the full description of the solute and solvent system is difficult in general because of the large number of degrees of freedom. The standard strategy to solve the problem is to combine, in some way, originally different theories or programs, which is the multi-physics approach.

In the multi-physics simulations, physical data are exchanged between separate program components, where we must transform not only formats but also their semantics, i.e., physical meanings of the data. In order to assist such dataexchanges with semantic transformations, we used a set of application program interfaces called Mediator (mediator-API) [11,1], which is included in the beta-version release of the NAREGI grid-middleware. Fig. 3(a) shows the total flow of this simulation, where the partial charge distribution of the solute and solvent molecules are exchanged each other through the mediator-API (Fig. 3(b)). In order to execute on the NAREGI grid, the flow is incorporated in the NAREGI Workflow tool (Fig. 3(c)).

In Fig. 4, we show results of this coupled calculation for methionine-enkephalin (75 atoms) and chignolin (138 atoms) in aqueous solution, where the partial charge distribution by water molecules are also shown around these molecules.

OpenFMO [6] is an open-licensed software platform to construct FMO applications under high-performance distributed computer environments. The current status of this development is in the end of Phase II. In Phase I, we introduced the OpenFMO framework and predicted a peta-scale performance on a hypothetical computer architecture [7]. In Phase II, we have tried to implement the skeleton by one-sided communications [12] under the PSI project [4]. In Phase III, we are going to extend the platform to the multi-physics simulations (see Fig. 5).

The main purpose of the Phase II in the development schedule of OpenFMO (Left of Fig. 5) was to correspond actual ex

…(Full text truncated)…

This content is AI-processed based on ArXiv data.