Self-scheduling parallel methods for multiple serial codes with application to WOPWOP

Abstract

This paper presents a scheme for efficiently running a large number of serial jobs on parallel computers. Two examples are given of computer programs that run relatively quickly, but often they must be run numerous times to obtain all the results needed. It is very common in science and engineering to have codes that are not massive computing challenges in themselves, but due to the number of instances that must be run, they do become large-scale computing problems. The two examples given here represent common problems in aerospace engineering: aerodynamic panel methods and aeroacoustic integral methods. The first example simply solves many systems of linear equations. This is representative of an aerodynamic panel code where someone would like to solve for numerous angles of attack. The complete code for this first example is included in the appendix so that it can be readily used by others as a template. The second example is an aeroacoustics code (WOPWOP) that solves the Ffowcs Williams-Hawkings equation to predict the far-field sound due to rotating blades. In this example, one quite often needs to compute the sound at numerous observer locations, hence parallelization is utilized to automate the noise computation for a large number of observers.

A Comparison of Ffowcs Williams-Hawkings Solvers for Airframe Noise Applications

“…Occasionally, one of the processors would take much longer than all of the others to complete. To circumvent this problem, the master-slave paradigm employed by Long and Brentner 17 was used. Here, one node does no work other than to tell requesting nodes what patch they should process.…”

Numerical simulation of the blast impact problem using the Direct Simulation Monte Carlo (DSMC) method

“…A very easy and obvious way to parallelize any Monte Carlo problem is to simultaneously run different ensembles on different computers and then average them. Such a program is called an 'embarrassingly parallel' program and it should ideally give a 100% speedup [26]. However, this approach is not viable for this problem because of the memory constraints.…”

Landing Gear Aerodynamic Noise Prediction Using Unstructured Grids

“…The sound directivity in the medium-and far-field requires the use of a parallel version of the FW-H post-processing program 8,28 . For each radii away from the landing gear, 72 observer locations are defined, so that a resolution of 5 degree angle is obtained.…”