GPU Large-Scale Nonlinear Programming

📢 Large-Scale Nonlinear Programming on GPUs: State-of-the-Art and Future Prospects
🎓 Presenter: Sungho Shin, ANL / MIT
📆 Date/Time: Thurs, Apr 11
🌐 Where: Online - Interactive chat and video conference
Congratulations to Sungho Shin as the winner of the AIChE CAST Division W. David Smith Jr. Graduate Publication Award.
AIChE Computing & Systems Technology Division webinar.
Bio
Sungho Shin is an incoming assistant professor at the Chemical Engineering Department of Massachusetts Institute of Technology and a postdoctoral researcher at the Mathematics and Computer Science Division at Argonne National Laboratory. He received his Ph.D. from the University of Wisconsin-Madison. He was a Summer intern at Los Alamos National Laboratory and Argonne National Laboratory. His research interests include model predictive control, optimization algorithms, and their applications to large-scale energy infrastructures (such as natural gas and power networks). He is the main developer of the nonlinear optimization solver MadNLP.jl and the automatic differentiation/algebraic modeling tool ExaModels.jl. He was the winner of the W. David Smith, Jr. Graduate Publication Award, AIChE Annual Meeting CAST Directors’ Student Presentation Award, IFAC ADCHEM Young Author Award, IFAC NMPC Young Author Award. He was a recipient of the Korea Presidential Science Fellowship, Kwanjeong Fellowship, and Grainger Wisconsin Distinguished Graduate Fellowship.
Chat Messages
From John Hedengren : We have about 5 minutes left for the presentation and then will have a brief Q+A. Please put your question in the chat window or unmute your microphone to ask.
From Emrullah ERTURK : 1- Can this developed NLP solver be used with JuMP? 2- Can we use this solver with the MIP solver to solve MINLP problems?
From Ashfaq Iftakher : Why Hybrid KKT systems seem to require less number of iterations that Lifted KKT? Is the degree of ill conditioning less in Hybrid KKT?
From Tianhao Liu : 1- How do the sparsity and problem scale in general nonlinear optimization (rather than special cases like ACOPF) affect the acceleration of GPU? 2- Can GPU (and cuDSS) still significantly outperform CPU in linear systems (e.g., for IPMs’ KKT system)? ...I mean general linear systems.
From John Hedengren : Junho Park will take over the Q+A moderation. I need to go start a class at BYU. Excellent presentation Sungho!
From Laurens Lueg : Can you comment on whether applying problem-level decomposition, ie. further decomposing the overall KKT system into partitions based on the problem structure (e.g. Schur complement decomposition) benefits the condensed IPM using CUDSS?
to just give the full sparse system to the GPU-enabled linear solver?
From Ashfaq Iftakher : Thanks Sungho for the detailed explanation. Excellent Presentation!!
From Emrullah ERTURK : Thanks for the presentation and answers.
From Laurens Lueg : Thank you!
From Tianhao Liu : Wonderful work! Thank you.

[@Anonimka338]

Hello, sorry for asking Lagrange multipliers here but I wanna clear my doubts . Can I just solve mutlivariable Optimizations using this ?
1. Solve problem standard way(like in equation constraints) so assuming both are active.
2. Solve problem with one Constrait being inactive using thay constant so it will be zero. Another constrait is active
3. Again but now vica versa of 2.
4. Both constraits are inactive.
And then which will be min and max will be my answer. Is that right ? Also whaf if I have just one constrait inequality. Looks like it will be only two steps
1. Solve like in equation constrait so its active
2. Solve like its nonactive so zero.
BUT if I get some values of x and y it MUST satisfy the inequality of constrait in both cases as Constrait was annulated because of That Constant by which its multiplied but its still our limitations. Please tell me if I am right, because I solved many Optimization problems using this method and it worked everytime .

[@apm]

These tutorials may help: apmonitor.com/me575/index.php/Main/KuhnTucker