<P1>

P4: Long Duration Aeroacoustic Simulations

Author: Dilhara Jayasundara , Advisor: James Baeder (Department of Aerospace Engineering)

Problem Statement Presentation

Abstract
The proposed project is to develop a new version of the acoustic code, ACUM-3, that has the capability to compute acoustic pressure time histories at different observer locations for large number of time steps. An important aspect of this version is to have time step parallelization so that the large number of time steps will not significantly affect the computational time. This can be done by using the binning approach instead of the current double-time loop. The other concern will be the massive amount of input data that has to be handled before and within the computation. Currently, input data is loaded using a separate file each surface that contain contain geometric and pressure data for each element at each time step. A CFD mesh contains about 20,000 to 50,000 elements based on the geometric and aerodynamic features of the surface. Having 360 time steps per revolution these input file sizes are of the order of 1GB and a thousand times that will essentially lead to memory issues. Therefore, this has to be optimized looking into the possibility of doing part of the computation inside the GPU and reducing the number of time steps per revolution without significantly compromising accuracy. A single GPU is unlikely to handle so much of data in its memory or to perform the computation in a reasonable amount of time. Therefore, it is necessary to look into the possibility of distributing the work load across multiple GPUs using MPI (Message Passing Interface). A high performance computing (HPC) cluster such as Deepthought2 can be used for running the code.

Final Presentation , Final Report