The focus of this work lies on implementational improvements and, in particular, node-level performance optimization of the simulation software ls1-mardyn. Through data structure improvements, SIMD vectorization and, especially, OpenMP parallelization, the world’s first simulation of 2*10
13 molecules at over 1 PFLOP/sec was enabled. To allow for long-range interactions, the Fast Multipole Method was introduced to ls1-mardyn. The algorithm was optimized for sequential, shared-memory, and distributed-memory execution on up to 32,768 MPI processes.
«
The focus of this work lies on implementational improvements and, in particular, node-level performance optimization of the simulation software ls1-mardyn. Through data structure improvements, SIMD vectorization and, especially, OpenMP parallelization, the world’s first simulation of 2*10
13 molecules at over 1 PFLOP/sec was enabled. To allow for long-range interactions, the Fast Multipole Method was introduced to ls1-mardyn. The algorithm was optimized for sequential, shared-memory, and distribu...
»