Parallel isosurface extraction for 3D data analysis workflows in distributed environments

Abstract

In this paper we discuss the issues related to the development of efficient parallel implementations of the Marching Cubes algorithm, one of the most used methods for isosurface extraction, which is a fundamental operation for 3D data analysis and visualization. We present three possible parallelization strategies and we outline the pros and cons of each of them, considering isosurface extraction as stand-alone operation or as part of a dynamic workflow. Our analysis shows that none of these implementations represents the most efficient solution for arbitrary situations. This is a major issue, because in many cases the quality of the service provided by a workflow depends on the possibility of selecting dynamically the operations to perform and, consequently, the more efficient basic building block for each stage. In this paper we present a set of guidelines that permits to achieve the highest performance for the extraction of isosurface in the most common situations, considering the characteristics of the data to process and of the workflow. These guidelines represent a suitable example to support the efficient configurations of workflows for 3D data processing in a dynamic and complex computing environment

CUDA accelerated molecular surface generation

“…To provide a fairer comparison, we experimentally determined that the maximum achievable speedup using all the cores of the CPU with a the Message Passing Interface (MPI) version of the sequential algorithm is of about 4.4. This is mainly because of the load balancing issue of the isosurface extraction operation, because a good balancing of active and non‐active cells is rather difficult, as discussed in .…”

Influences between Performance Based Scheduling and Service Level Agreements

“…For the first class of applications, we selected the well-known High Performance Linpack (HPL) benchmark [15]. For the second, we realised a lightweight version of the application [16], characterized by a reduced computational cost, but still capable to maintain a representative run of the real application (ISO). A deep discussion about the definition and effectiveness of a two-level benchmark methodology has been presented in [17].…”