Brieuc lehmann

Have used `memory-profiler` to check RAM usage (see updated script above). Oddly, there doesn't appear to be too much of a difference with and without caching, but I'm not 100%...

I simplified the code slightly (see edits above) so that it only profiles the `ts.genetic_relatedness` call. Here are the results with `n_samples = 500` across 3 runs of the same...

Thanks @petrelharp ! Agreed with your point about combining `update_node_summary` and `update_running_sum` - is it worth doing this now or shall I wait until we're sure we want to include...

And here the memory-profiler results for `n_samples = 2000`, TL;DR largely the same again With caching: ``` Line # Mem usage Increment Occurences Line Contents ============================================================ 9 111.1 MiB 111.1...

Ah OK, now with `n_samples = 500` and `recombination_rate = 1e-7` corresponding to ~250K trees and ~150k nodes. We now see a slight improvement without caching. With caching: ``` Line...

Just been playing around with this and it seems like there might be some numerical stability issues. I tried @petrelharp 's code above with different seeds and occasionally get an...

OK, still some bizarre behaviour on rescomp. See the following code to perform matrix-vector multiplication using the same tree sequence as above. I am using the old code for comparison...

I'm also getting the problem when I run the above code in the command line via a script... Never used valgrind before but could look into it if that would...

Yep, I'd be happy to! Please could you point me towards a relevant function/method that I could use as a template?

Ping @petrelharp - I'd be happy to write some tests and more documentation in the coming days so that we can get this merged, but I think I'm not able...