Communication Reducing Algorithms for Distributed Hierarchical N-Body Problems with Boundary Distributions

Abstract

Reduction of communication and efficient partitioning are key issues for achieving scalability in hierarchical N -Body algorithms like FMM. In the present work, we propose four independent strategies to improve partitioning and reduce communication. First of all, we show that the conventional wisdom of using space-filling curve partitioning may not work well for boundary integral problems, which constitute about 50% of FMM's application user base. We propose an alternative method which modifies orthogonal recursive bisection to solve the cell-partition misalignment that has kept it from scaling previously. Secondly, we optimize the granularity of communication to find the optimal balance between a bulk-synchronous collective communication of the local essential tree and an RDMA per task per cell. Finally, we take the dynamic sparse data exchange proposed by Hoefler et al. [1] and extend it to a hierarchical sparse data exchange, which is demonstrated at scale to be faster than the MPI library's MPI Alltoallv that is commonly used.

Optimizations of Unstructured Aerodynamics Computations for Many-core Architectures

“…Scalability of scientific workloads on next-generation computing systems is critically dependent upon the performance of a shared-memory multi and many-core compute node [82], to which this paper adds a chapter for the important problem class of unstructured meshes.…”

Nonlinear time‐domain wave‐structure interaction: A parallel fast integral equation approach

“…This parameter determines the number of levels to the octree created for the FMM; a small N crit value will result in more levels, whereas a large value will result in fewer. One can see typical values for different distributions of points in Abdul-Jabbar et al 82 For any given box, we store the center, size, and either the list of "child" boxes or collocation nodes or elements stored.…”

FFT, FMM, and multigrid on the road to exascale: Performance challenges and opportunities

“…In this study, we adopt a tree structure that is similar to the one described in [10,11] where FMM uses a separate tree structure for the local and global trees. Each leaf of the global tree is a root of a local tree for a particular MPI process.…”