Hello,
This ticket is due to concerns regarding the performance of CarpetX running with CPUs in the new machine MareNostrum 5.

The first issue I had was the memory needed to run a q=1 configuration, coming across several out-of-memory errors (regardless of the number of nodes requested). The standard nodes in MN5 have 2GB per core (112 cores per node), so I had to switch to the high memory ones to be able to make it run (this is the line `#SBATCH --constraint=highmem` in the attached batch file). Is this also the case for other machines?

The second issue is that now that is running (ncells=128 and 1 node) it’s going awfully slow ~0.07 M/hr, less than I would expect from using one node. These type of binary configurations we have been running them in 2~4 nodes reaching speeds of around 4~7 M/hr with Carpet and at higher resolution.

Attached you find the parfile, batch file, and output.

attachment:	q1.__0._0._0.8__0._0._0.8__e0.5.par
attachment:	output.log
attachment:	run_cactus.sh

--
Ticket URL: https://bitbucket.org/einsteintoolkit/tickets/issues/2882/running-bbh-with-carpetx-with-cpus-high