Improvement Cache Efficiency of Explicit Finite Element Procedure and its Application to Parallel Casting Solidification Simulation

TL;DR

This paper introduces a cache-efficient finite element method for casting solidification simulation, achieving significant performance improvements through data blocking, mesh reordering, and parallel algorithms on complex geometries.

Contribution

It presents a novel cache-efficient finite element algorithm with mesh reordering and a high-quality mesh decomposition method for improved simulation performance.

Findings

10-20% performance improvement from mesh reordering

1.2-2.2x speedup with the cache-efficient algorithm

Nearly linear speedup of the parallel solver on a Linux cluster

Abstract

A simple method for improving cache efficiency of serial and parallel explicit finite procedure with application to casting solidification simulation over three-dimensional complex geometries is presented. The method is based on division of the global data to smaller blocks and treating each block independently from others at each time step. A novel parallel finite element algorithm for non-overlapped element-base decomposed domain is presented for implementation of serial and parallel version of the presented method. Effect of mesh reordering on the efficiency is also investigated. A simple algorithm is presented for high quality decomposition of decoupled global mesh. Our result shows 10-20 \% performance improvement by mesh reordering and 1.2-2.2 speedup with application of the presented cache efficient algorithm (for serial and parallel versions). Also the presented parallel solver (without cache-efficient feature) shows nearly linear speedup on the traditional Ethernet networked Linux cluster.