R.C.Giles and P.Tamayo, Proc of SHPCC'92, IEEE Computer Society (1992), p. 240.  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... scripting languages, visualization, parallel computing, SPaSM, SWIG Introduction With the development of massively parallel supercomputers, the materials science community has experienced an unprecedented explosion in both the size and complexity of short range molecular dynamics simulations [1,2,3,4,5]. The method of molecular dynamics (MD) has been used since the 1950 s for a variety of computational problems in physics, chemistry, and materials science [6] The idea is really quite simple given a collection of atoms, we solve Newton s equation of motion F=ma and track the atoms ....

R.C.Giles and P.Tamayo, Proc of SHPCC'92, IEEE Computer Society (1992), p. 240.

....[7, 8] which have the potential for multi million particle simulations. In fact, recent work by Holian and collaborators [9, 10] have demonstrated 2D simulations for 1:6 Theta 10 7 atoms on a 64K processor CM 200. Work on MIMD implementaions of large scale MD is now also starting to appear [11, 12]. In this paper we present a scalable parallel MD algorithm which allows for the simulation of at least 10 8 particles interacting via a relative short range potential. We have implemented this algorithm on the Connection Machine 5 (CM 5) from Thinking Machines Corporation (TMC) and thus ....

....will simply move to a cell on the same processor. This fact, allow us to update the data structures for a very large system of particles very quickly. The redistribution algorithm rebuilds all necessary data structures. There is no searching for neighbors or the construction of neighbor lists [11]. There is also an advantage in that this procedure is called after every time step so we are assured that the data is in the proper layout before each force calculation. As mentioned earlier, the redistribution procedure tends to fragment data. Deleted particles are generally stored together with ....

R. C. Giles and P. Tamayo, Proc. of SHPCC'92, IEEE Computer Society (1992), p. 240.

....However, it is parallel multicomputers that seem to hold the greatest potential for reaching beyond 10 9 atoms. Coarse grained transputer systems have been used in successful 2D simulations with 10 6 atoms [6] Data parallel SIMD implementations on the CM 2 have also recently been presented [7, 8] which have the potential for multi million particle simulations. In fact, recent work by Holian and collaborators [9, 10] have demonstrated 2D simulations for 1:6 Theta 10 7 atoms on a 64K processor CM 200. Work on MIMD implementaions of large scale MD is now also starting to appear [11, 12] ....

P. Tamayo, J. P. Mesirov, and B. M. Boghosian, Proc. of Supercomputing 91, IEEE Computer Society (1991), p. 462.

....provides large amounts of cost effective memory that is rarely available on any other type of supercomputer systems. Large memory systems with tens of GBytes allow for simulations of a size that was not practical in the past. For example in materials science, the method of Molecular Dynamics (MD) [1, 2] running on MPPs is currently being used to study fracture and crack propagation on length scales that was not possible just a few years ago. In order to make reasonable comparisons with experimental data it is often necessary to be able to simulate features on a micron scale. Realistic 3D MD ....

P. Tamayo, J.P. Mesirov and B.M. Boghosian, Proc. of Supercomputing 91, IEEE Computer Society, 1991, pp: 462.

.... of massively parallel supercomputers has generated considerable interest in developing fast parallel algorithms for performing multi million atom MD simulations [2, 3, 4, 5, 6, 7, 8] On state of the art MPP systems, simulations with more than 100 million (10 8 ) atoms can now be performed[9]. Simulations of this size will be crucial in performing realistic experiments in materials science where it will be necessary to simulate hundreds of millions or even billions of atoms to realistically capture the behavior of dislocations, fracture, and crack propagation. We have developed a ....

....modifications to achieve high performance on the CM5. This has improved the time required to perform a single timestep by roughly a factor of ten. In addition, SPaSM was one of the winners in the 1993 IEEE Gordon Bell prize competition for achieving a speed of 50 Gflops on a 1024 processor CM 5[9]. In this paper, we provide a brief overview of our general MD algorithm and focus on the enhancements that have allowed us to achieve high performance on the CM 5. We also present recent timings. Our code has been implemented in ANSI C with explicit calls to the CMMD message passing library 1 . ....

[Article contains additional citation context not shown here]

P. S. Lomdahl, P. Tamayo, N. Grønbech-Jensen, and D. M. Beazley, Proc. of Supercomputing 93, IEEE Computer Society (1993), p. 520-527.

.... most MD simulations have been limited to small simulations involving less than a million atoms[2, 3, 10, 11] However, the development of massively parallel supercomputers has generated considerable interest in developing fast parallel algorithms for performing multi million atom MD simulations [2, 3, 4, 5, 6, 7, 8]. On state of the art MPP systems, simulations with more than 100 million (10 8 ) atoms can now be performed[9] Simulations of this size will be crucial in performing realistic experiments in materials science where it will be necessary to simulate hundreds of millions or even billions of atoms ....

R. C. Giles and P. Tamayo, Proc. of SHPCC'92, IEEE Computer Society (1992), p. 240.

....nearby. This substantially reduces the complexity of the problem, but there are still many computational difficulties associated with short range MD simulations. Due to limitations in computing resources, most MD simulations have been limited to small simulations involving less than a million atoms[2, 3, 10, 11]. However, the development of massively parallel supercomputers has generated considerable interest in developing fast parallel algorithms for performing multi million atom MD simulations [2, 3, 4, 5, 6, 7, 8] On state of the art MPP systems, simulations with more than 100 million (10 8 ) atoms ....

.... most MD simulations have been limited to small simulations involving less than a million atoms[2, 3, 10, 11] However, the development of massively parallel supercomputers has generated considerable interest in developing fast parallel algorithms for performing multi million atom MD simulations [2, 3, 4, 5, 6, 7, 8]. On state of the art MPP systems, simulations with more than 100 million (10 8 ) atoms can now be performed[9] Simulations of this size will be crucial in performing realistic experiments in materials science where it will be necessary to simulate hundreds of millions or even billions of atoms ....

[Article contains additional citation context not shown here]

P. Tamayo, J. P. Mesirov, and B. M. Boghosian, Proc. of Supercomputing 91, IEEE Computer Society (1991), p. 462.

....by the MD method, which is typically tens or maybe hundred of nano sec. at best. Ideally one would like to use the MD method for second long simulations with at least 10 9 atoms. While this goal is still very far away, there is substantial current interest in the development of fast MD algorithms[2, 3, 4, 5, 6, 7] which allow for the simulation of at least million atom systems. We have developed an new scalable MD algorithm based on a message passing multi cell approach which allows for simulating at least 10 8 particles interacting via a relative short range potential. We have implemented the algorithm ....

R. C. Giles and P. Tamayo, Proc. of SHPCC'92, IEEE Computer Society (1992), p. 240.

....by the MD method, which is typically tens or maybe hundred of nano sec. at best. Ideally one would like to use the MD method for second long simulations with at least 10 9 atoms. While this goal is still very far away, there is substantial current interest in the development of fast MD algorithms[2, 3, 4, 5, 6, 7] which allow for the simulation of at least million atom systems. We have developed an new scalable MD algorithm based on a message passing multi cell approach which allows for simulating at least 10 8 particles interacting via a relative short range potential. We have implemented the algorithm ....

P. Tamayo, J. P. Mesirov, and B. M. Boghosian, Proc. of Supercomputing 91, IEEE Computer Society (1991), p. 462.

....MD simulations of this size requires at least hundreds of millions of atoms, preferably more. In this paper, we describe recent advances in the development of our MD code for MPPs, SPaSM. Our basic algorithm has been described in details elsewhere [3, 4] as well as performance and scaling results [5]. Here we present our experience over the past year with performance optimization and code portability on a variety of MPP systems. While specific to our MD code, we believe that the lessons learned can be applied to any major scientific code on MPPs. The outline of the papers is as follows: in ....

....of 128 Mflops. Given that our C code ran exclusively on the SPARC, it was clear that getting high performance would depend on utilizing the vector units. To use the vector units, we rewrote our force calculation in CDPEAC, a set of C macros for programming in the VU assembler language DPEAC [5]. In addition, we developed a special memory caching scheme, also written in CDPEAC, to reduce the number of memory accesses [1] As a result of these modifications, we were able to achieve calculation rates between 25 and 50 Mflops node[5] Unfortunately, we have experienced a wide range of ....

[Article contains additional citation context not shown here]

P.S. Lomdahl, P. Tamayo, N. Grønbech-Jensen and D.M. Beazley, Proc. of Supercomputing 93, IEEE Computer Society (1993), pp: 520--527.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC