NCSI: GPU-accelerated Sub-grid Models for Multi-phase Fluid Dynamics

Status	Completed
Mentor Name	Spencer Bryngelson
Mentor's XSEDE Affiliation	Research Allocation
Mentor Has Been in XSEDE Community	1-2 years
Project Title	GPU-accelerated Sub-grid Models for Multi-phase Fluid Dynamics
Summary	Sub-grid models for flowing dispersions are essential for the high-fidelity simulations of cough droplets leaving your mouth, bubble clouds cavitating near ship propellers, and more. Still, these sub-grid models are complex with high arithmetic intensity, dominating the cost of their associated simulations. We will develop a GPU-accelerated implementation of these models to enable next-generation multi-phase flow simulation.
Job Description	The student will implement sub-grid physics models into MFC, a hybrid MPI+OpenACC open-source multi-phase flow solver maintained, in part, by the mentor. This will start with implementation of a simple model entailing basic bubble dynamics, validation of correctness, and profiling. As time allows, more realistic models for the dynamics of flowing bubbles, droplets, and solid particles will be added. The speed-up relative to a CPU baseline will be established.
Computational Resources	The student will use Bridges2 and Expanse GPU nodes. The mentor has already utilized these resources and confirmed their suitability for these simulations. MFC, our code, currently has no trouble running utilizing these resources. A relatively low number of node hours will be required as the primary challenge lies in the implementation, with MFC already performing well on weak- and strong-scaling metrics.
Contribution to Community	This project will utilize the latest XSEDE resources, including full multi-node multi-GPU simulations via MPI+OpenACC and Nvidia MPS. This will test their capabilities for large multi-phase flow simulations that are becoming ever more important in our world. For example, the Bell Prize this year went to COVID-related work, which our project closely brushes up against, albeit from a different angle. The success of this work will also showcase the capabilities of GPUs for accelerating sub-grid multi-phase fluid dynamics models. This is currently an under-utilized tool in the multi-phase fluid mechanics community and will draw attention to it and associated XSEDE resources.
Position Type	Intern
Training Plan	The mentor will play a semi-hands-on role with the student, bringing them up-to-speed on the latest best practices in GPU computing, OpenACC, MPI, and more as appropriate. In particular, the code will be fully open source and maintained in such a fashion to allow for synergistic communication with other Bryngelson Research Group members.
Student Prerequisites/Conditions/Qualifications	Experience with compiled programming languages (Fortran or C preferred). Some knowledge of distributed memory programming interfaces like MPI are preferred.
Duration	Semester
Start Date	08/01/2021
End Date	11/01/2021