Overestimation of the number of links
I am running an EAGLE-200 box with both hydro and gravity. It crashed because it ran out of memory. I have restarted it with more information. One of the culprit is the number of foreign part/gpart we allocated (see #456 (closed)) but the more interesting one is the number of link "objects".
On this run, we have e->size_links == 4'568'936'336 (!!!) but e->nr_links == 1'145'103.
This means that the estimate of the number of links is wrong by a factor 4000. It also means that for this run we are allocating more than 64GB per MPI rank for the links!
The obvious fix is to come with a better estimate of the number of links we need.
Interesting numbers for one rank are:
- Number of top-level cells: 262'144 (64^3)
- Number of cells: 24'644'944
- Number of tasks: 1'743'339