Tinkertoy parallel programming: a case study with Zoltan

Abstract

Abstract-As the need for complex parallel simulation software grows, better strategies for efficient and effective software development become important. We advocate a toolkit -or "tinkertoy" -approach to parallel application development. By providing efficient implementations of basic services commonly needed by applications, toolkits allow application developers to benefit from others' research, compare algorithms, and save time for their own development. Unlike large frameworks, toolkits provide these services with light-weight interfaces and little or no restriction on application data structures, making them easy to use in both new and existing applications. In this paper, we describe features of effective toolkit design, using the Zoltan parallel, dynamic data management toolkit as an example.

Reordering and incomplete preconditioning in serial and parallel adaptive mesh refinement and coarsening flow solutions

“…The recursive local nature of AMR/C impacts solver selection and solver performance. Another package that provides parallel services for dynamic, unstructured, and/or adaptive finite element simulations is Zoltan toolkit [29,30]. Because the changes in the mesh are local, introducing or 805 removing a node (grid point) will change the associated graph locally, and this has implications regarding reordering strategies and solver choices [27].…”

“…In this way, libMesh has proved a valuable testbed for a wide range of physical applications. Another package that provides parallel services for dynamic, unstructured, and/or adaptive finite element simulations is Zoltan toolkit [29,30]. Zoltan, released recently as part of Trilinos [3], includes only versions of the nested dissection matrix ordering algorithm, whereas PETSc (used by libMesh) provides several ordering algorithms.…”

Dynamic load balancing for large-scale adaptive finite element computation

“…、递归惯量二分方法 RIB (Recursive Inertial Bisection) [24]、空间填充曲线划分方法 SFC (Space-Filling Curve) [33,12,26] 等. 基于图的划分方法是 利用对象的拓扑信息, 如网格中单元的邻居关系, 常用方法有递归图二分方法 (Recursive Graph Bisection)、贪 婪算法、递归谱二分方法 RSB (Recursive Spectral Bisection) [24,25]、K-L 算法 (Keinighnan-Lin Algorithm)、 多水平方法 (multilevel methods) [18,19,14,15,17] 和扩散方法 (diffusive methods) [20,21] 等. 图方法和几何 方法各有优缺点.…”