---

Parallel Mesh Generation Using COTS Software

 

Nikos Chrisochoides

 

Published in Keynote talk at the Annual Computational Science and Engineering Research Symposium, University of I, April, 2005

 

Abstract

 

Existing parallel mesh generation codes are based on the parallelization of well known sequential mesh generation technology. Given that it takes a very long time to develop the software infrastructure for sequential industrial-strength mesh generation libraries, it is clear that traditional parallelization approaches deliver technology that is outdated. This problem becomes more serious if one considers that improvements of sequential codes in terms of quality, speed, and functionality are open ended. In this talk we present a COTS (commercial of-the-shelf) based approach to parallel mesh generation for addressing this serious problem. We will discuss our experience from three parallel meshing methods using public state-of-the-art sequential software for both 2-D (Triangle from CMU) and 3-D (SolidMesh from MSU) geometries. Parallel mesh generation procedures decompose the original mesh generation problem into N smaller subproblems that can be meshed in parallel using P (<< N) nodes. The subproblems can be formulated to be either tightly or partially coupled or even decoupled to each other. The coupling of the subproblems determines the intensity of the communication and the degree of dependency (or synchronization) between the subproblems. In this talk we will overview some of our work on tightly-, partially-coupled, and decoupled methods based on Delaunay and advancing front techniques. We will also discuss the lessons we learned. Our parallel mesh generation codes are implemented on top of a runtime system (PREMA) that provides support for one-sided communication, remote service request, global address space in the context of data mobility, transparent routing of messages, and automatic dynamic load balancing. We will briefly describe PREMA and present experimental data that indicate performance improvements of 30% to 50% over plain MPI codes for generating tetrahedral meshes that vary from 100 million to 1.2 billion elements. We will conclude the talk with open problems and future directions. This work was supported in part by the following NSF grants: Career Award #CCR-0049086, ITR #ACI-0085969, RI #EIA-9972853, NGS #EIA-0203974, and ITR-0312980.