Professor David Keyes's Lectures

Exaflop/s, Seriously! PUBLIC LECTURE
Wednesday, February 16, 2011 - 4:00 p.m., Alumni Center Ballroom (refreshments will be served at 3:30 p.m.)

Abstract: The paths to the exascale summit are debated, but all are narrow and treacherous, constrained by physics, cost, power consumption, programmability, and reliability. Nevertheless, computational scientists have an insatiable appetite for extreme performance arising from requirements of resolution, fidelity, dimension, artificial boundaries, multiphysics, inversion and assimilation, optimization and control, and uncertainty quantification. Looking beyond the Pax MPI of the past two decades, we propose architectural, programming model, and algorithmic directions in one of the great globally-joined scientific quests of the next decades.

Applications that Break Techniques ENGINEERING LECTURE
Tuesday, February 15, 2011 - 4:15 p.m., Engineering Lecture, A226 College of Engineering, Building A (refreshments will be served at 4:00 p.m.)

Abstract: Following the rapid colonization of any new science, advances are driven mainly by what does not fit and it is the latter that attracts scientific attention. The best news for an algorithmicist is that a method or a piece of software breaks. Therein, with any providence, lies a fruitful collaboration with a scientist or engineer. In this talk, we revisit a few favorite occasions when something broke and we muse on some transferable principles of how to proceed. Audience participation is required in order to properly conclude the talk.

A Nonlinearly Implicit Manifesto MATHEMATICS LECTURE
Thursday, February 17, 2011 - 3:30 p.m., Mathematics Lecture, 101 Love Building (refreshments will be served at 3:00 p.m. in 204 Love Building)

Abstract: Many simulations must be followed over time intervals long compared to other timescales in the system. Often, the phenomena associated with the shortest timescales may be assumed to be in equilibrium relative to dynamics of interest, but they control the computational timestep if explicit integration is used, so that even weak computational scaling cannot be achieved. A high-order timestepping scheme would permit large timesteps for computational economy; however, operator-splitting thwarts this objective. The challenges of uncertainty quantification, inverse problems, and multiphysics coupling, like the above, are most naturally tackled with fully nonlinearly implicit formulations. We illustrate the issues and advocate the nested algorithmic framework of Newton-Krylov-Schwarz.