Experience with Distributed Memory Delaunay-based Image-to-Mesh Conversion Implementation

TL;DR

This study evaluates the scalability and efficiency of distributed memory parallel algorithms for 3D image-to-mesh conversion using Delaunay triangulation, highlighting performance trade-offs between shared and distributed memory implementations.

Contribution

It introduces a distributed memory implementation of a Delaunay-based image-to-mesh algorithm and compares its performance with shared memory approaches, revealing scalability challenges.

Findings

Distributed implementation achieves up to 5.4x speedup with 180 cores.

MPI-MW implementation has higher overheads, making it slower than shared memory on the same cores.

Shared memory implementation is more efficient than distributed memory for the tested configurations.

Abstract

This paper presents some of our findings on the scalability of parallel 3D mesh generation on distributed memory machines. The primary objective of this study was to evaluate a distributed memory approach for implementing a 3D parallel Delaunay-based algorithm that converts images to meshes by leveraging an efficient shared memory implementation. The secondary objective was to evaluate the effectiveness of labor (i.e., reduce development time) while introducing minimal overheads to maintain the parallel efficiency of the end-product i.e., distributed implementation. The distributed algorithm utilizes two existing and independently developed parallel Delaunay-based methods: (1) a fine-grained method that employs multi-threading and speculative execution on shared memory nodes and (2) a loosely coupled Delaunay-refinement framework for multi-node platforms. The shared memory implementation uses a FIFO work-sharing scheme for thread scheduling, while the distributed memory implementation utilizes the MPI and the Master-Worker (MW) model. The findings from the specific MPI-MW implementation we tested suggest that the execution on (1) 40 cores not necessary in the same single node is 2.3 times faster than the execution on ten cores, (2) the best speedup is 5.4 with 180 cores again the comparison is with the best performance on ten cores. A closer look at the performance of distributed memory and shared memory implementation executing on a single node (40 cores) suggest that the overheads introduced in the MPI-MW implementation are high and render the MPI-MW implementation 4 times slower than the shared memory code using the same number of cores. These findings raise several questions on the potential scalability of a "black box" approach, i.e., re-using a code designed to execute efficiently on shared memory machines without considering its potential use in a distributed memory setting.