Present: Erik, Roland, Yosef, Helvi, Swapnil, Liwei, Lucas, Cheng-Hsin

Performance of Z4cow using Carpet vs CarpetX ======= Liwei compared the performance of Z4cow between Carpet and CarpetX versions, and found that the CarpetX version is slower than Carpet by 30% to 50%. - Setup: 1 AMR level, with subcycling, 128^3 grid on 1 MPI process, 32 time steps. - Erik suggested that timing information is needed to see which part is the most expensive. - Liwei is thinking about profiling using HPCToolkit, but will try using the Carpet timing information first. - Roland asked if Z4cow has been compared with LazEv in terms of performance, but this hasn't been done yet. - Liwei was able to re-run the tests and show the timing information at the end of the meeting - Erik pointed out that CarpetX does domain decomposition by default into 64 boxes whereas Carpet does not, so the CarpetX test will need to use a grid size of 128 for an even comparison with Carpet. - Erik suggested that the number of RHS calls could be different and that Liwei should check it with a counter. - Erik mentioned that the AMReX tile sizes affect the performance a lot, so Liwei should also experiment with using different tile sizes to get the best performance from CarpetX.

Fix for NewRadX when symmetries are applied ====== - Roland asked about the status of the pull requests for the NewRadX fix. - Cheng-Hsin reported that the pull request under CarpetX has been reviewed but still awaits formal approval from Erik before it can be merged. - The pull request under SpacetimeX needs a reviewer, and Lucas volunteered.

Subcycling paper ====== - Awaiting reviewing by co-authors, and there is still some text in the discussion that needs to be written up. Will ask people if they would like to be listed as co-authors or left out.

Updates from Lucas ====== - Lucas is looking into dereferencing issues when using function pointers, and is also looking into calling functions directly instead. - Lucas, Steve, and Max (PhD student) have been testing hybrid Runge-Kutta integrators on BBH simulations, where some of the intermediate RK steps are replaced with the RHS evaluated at previous timesteps, making it a multi-step method. - The hybid method maintains the accuracy and maintains the RHS evaluated in previous steps, but the downside is that more memory may be required. - Comparing the BBH waveform using standard RK4, 2-step hybrid, 3-step hybrid methods, the error is small (maximum 1e-9), and the hybrid method runs much faster than standard RK4. - The code is in the branch feature/RKAB in the CarpetX repo, and Lucas showed how to add the hybrid RK method to an evolution thorn by adding tags to the state variables.