The Cyber-ShARE Center of Excellence would like to invite you to an invited lecture by Dr. Carsten Burstedde from the Institute for Computational Engineering and Sciences at The University of Texas at Austin. See you there!
Location and Time: Classroom Building, room 402, Tuesday, April 12, 2:30 – 3:30
Many scientific processes, such as earth’s mantle convection or ice sheet dynamics, can be efficiently simulated using parallel adaptive mesh refinement (AMR). This method assigns an individual section of the computational grid to each processor. A central algorithm used in the distribution of grids is MPI_Allgather, which allows each processor to share its number of grid cells with every other processor, and to receive every other’s number of grid cells in return. While a naive implementation of MPI_Allgather performs badly on large numbers of processors, a recursive divide-and conquer algorithm works well up to the full size of contemporary supercomputers. In this talk, I will develop such a scalable MPI_Allgather algorithm step by step, and comment on some dos and don’ts for parallel algorithm development.