High-Order for the Price of Low-Order: Towards GPU Accelerated Flow Solvers for Exascale Applications

Event Sponsor: 
Mathmatics and Computer Science Division Seminar
Start Date: 
Feb 22 2018 - 10:30am
Building/Room: 
Building 240/Room 4301
Location: 
Argonne National Laboratory
Speaker(s): 
Noel Chalmers
Speaker(s) Title: 
Virginia Tech
Host: 
Misun Min

In this talk I will discuss the technical and design challenges of developing high-order finite element flow solvers which can efficiently take advantage of GPU accelerators present in current and next-generation Leadership Computing Facilities. I will detail how despite high-order methods being well-suited for approximating smooth flows, the GPU performance of operators and linear solves degrades as the polynomial degree increases. I will describe how we can construct empirically determined roof-line models to guide the optimization of GPU compute kernels in order to maximize performance, and how choosing a basis which gives sparser operators can yield even greater performance gains. I will then discuss performance of GPU accelerated preconditioners for high-order elliptic finite element problems and describe a new low-order preconditioning technique for triangle elements where the low-order and high-order interpolation nodes do not necessarily coincide, but are connected via a least squares projection. I will present some computational results obtained in three dimensional incompressible flows and scaling studies.
 

Miscellaneous Information: 

BlueJeans Meeting URL: https://bluejeans.com/150337078/1163 / Meeting ID: 150 337 078 / Participant Passcode: 1163

Please click here [schedule.ics] to add this event to your calendar.