Date of Award

Fall 12-2020

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Committee Director

Nicolaos Chrisochoides

Committee Member

Andrey Chernikov

Committee Member

Michael Park

Committee Member

Ravi Mukkamala

Abstract

Preliminary results of a long-term project entailing the parallelization of an industrial strength sequential mesh generator, called Advancing Front Local Reconnection (AFLR), are presented. AFLR has been under development for the last 25 years at the NSF/ERC center at Mississippi State University. The parallel procedure that is presented is called Pseudo-constrained (PsC) Parallel Data Refinement (PDR) and consists of the following steps: (i) use an octree data-decomposition scheme to divide the original geometry into subdomains (octree leaves), (ii) refine each subdomain with the proper adjustments of its neighbors using the given refinement code, and (iii) combine all subdomain data into a single, conforming mesh. Parallelism was achieved by implementing Pseudo-constrained Parallel Data Refinement AFLR (PsC.AFLR) on top of a runtime system called Parallel Runtime Environment for Multi-computer Applications (PREMA). During run time, the PsC.AFLR method exposes data decomposition information (number of subdomains waiting to be refined) to the underlying runtime system. In turn, this system facilitates work-load balancing and guides the program’s execution towards the most efficient utilization of hardware resources. Preliminary results, on the mesh refinement operation, show that the end-user productivity (measured in terms of elements refined per second) increases as the number of cores in use are increased. When using approximately 16 cores, PsC.AFLR outperforms the serial AFLR code by about 11 times. PsC.AFLR also maintains its stability by generating meshes of comparable quality. Although it offers good end-user productivity, PsC.AFLR suffers in its capability to generate meshes with the same level of density or quality as that of the serial AFLR software due to the constraints set by subdomain boundaries that are required to successfully execute AFLR. These constraints demonstrate that it is not ideal to use AFLR in a black box manner when parallelizing the software. Its source code must be modified to a non-trivial extent if one wishes to remove these constraints and maximize the end-user productivity and potential scalability.

DOI

10.25777/appr-3169

ISBN

9798557051477

Share

COinS