search

IDAS-Durham / JUNE

June is a framework for agent based modelling in an epidemiological and geographical context.
GNU General Public License v3.0
41 stars 11 forks source link

parallelisation by NHS region #282

Closed bnlawrence closed 3 years ago

bnlawrence commented 4 years ago

We have discussed the possibility of using node level parallelisation for regions, and core-level parallelisation for loop.

Probably the easiest regional decomposition to start with would be to look at running each of the NHS regions in parallel and moving people between them as required at the beginning or end of each timestep.

To do that, we need to understand a bit more about the geolocality in the model. Where and how are people moved, and how is their geographical locations used?

With respect to the second question, naively I assume there is some convolution of infection probability and people "met" in the Interaction part of the model. Is that right?

But the first question is the most pressing.

bnlawrence commented 4 years ago

I have moved my discussion of performance to the new issue above.

florpi commented 4 years ago

I'll write one last closing message on this issue to show that the parallel code is giving the same results with the same infection seed, as the parallel code. As mentioned earlier, I haven't fixed the random seed (only for the infection seed, so the initial infected people are in fact the same), thus I have taken the mean of 20 realisations to make this comparison,

image

sadielbartholomew commented 4 years ago

Don't worry about adding the odd comment, all, my warning is more about if we continued for another week to comment at a similar rate to that of the past week, we might start to run into trouble...

I love that plot, BTW, @florpi! It's great that the sequential & current parallel code are giving the same results.