deploy: 656aab7

Author: tlunet

events/14th-pint-workshop/index.html

@@ -227,7 +227,7 @@ <h4 id="main-actions">Main Actions</h4>
 <h4 id="full-summary">Full Summary</h4>
 
 <ul>
-  <li>Redundant computation (<a href="https://crd.lbl.gov/divisions/amcr/computational-science-dept/anag/about/staff-and-postdocs/hans-johansen/">Hans</a> to summarize).</li>
+  <li><a href="https://crd.lbl.gov/divisions/amcr/computational-science-dept/anag/about/staff-and-postdocs/hans-johansen/">Hans</a> asked about removing serial bottlenecks for any coarse operators. In other fields, redundant computation is used to remove latency and improve scaling. In (space) MG methods consolidation is another technique. <a href="https://github.com/JHopeCollins">Josh</a> suggested wavefront-type methods, and <a href="https://people.llnl.gov/falgout2">Rob</a> commented this is just changing the constant for an already O(N) method. There was a suggestion to develop performance models like what <a href="https://github.com/danielru">Daniel</a> had presented in his talk.</li>
   <li><a href="https://www.unige.ch/~gander">Martin</a> cautioned against the use of the word “optimal” without more specific information - this comes from the DD community where it really means “mesh independent scaling”.</li>
   <li>Does PinT work for hyperbolic problems? See <a href="https://www.cambridge.org/core/services/aop-cambridge-core/content/view/E580B25A44766E6729A65D7AF05B1198/S0962492924000072a.pdf/time_parallelization_for_hyperbolic_and_parabolic_problems.pdf">Martin’s Acta Numerica paper</a> - there are many methods that are not multilevel that can work. Multilevel is a challenge.</li>
   <li><a href="https://github.com/JHopeCollins">Josh</a> pointed out a common subproblem shared by ParaDiag and REXI methods: solution of systems that look like backwards Euler steps with complex timestep (which <a href="https://github.com/colinjcotter">Colin</a> clarified lead to an indefinite-Helmholtz-like problem on elimination to one variable).</li>