Add initial checkpoint and restore functionality.

[plietar]

This allows a user to run a simulation for a number of time steps and save the state of the simulation. The state can then be restored to continue the simulation.

This is,for example, useful to compare the impact of an intervention based on different parameters. The simulation can be run up to the point of an intervention, when the parameters' value don't matter. The simulation can then be resumed multiple times but with different intervention parameters each time.

For now, only variables are supported. Events will be added next.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (3800945) 96.28% compared to head (a329052) 96.37%.

Additional details and impacted files

@@            Coverage Diff             @@
##              dev     #178      +/-   ##
==========================================
+ Coverage   96.28%   96.37%   +0.08%     
==========================================
  Files          36       36              
  Lines        1722     1764      +42     
==========================================
+ Hits         1658     1700      +42     
  Misses         64       64              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[plietar]

I still need to add some unit tests to make sure the checkpoint/restore works correctly DONE, but I'm interested in getting some feedback in the meantime.

The long term goal is to support mrc-ide/malariasimulation#196. There's obviously still a long way to go for this, especially since malariasimulation maintains quite a lot of state that isn't visible to individual, but this is the first step in that direction.

[giovannic]

Very minor suggestions, I love the way this is done.

I'm curious about whether users would want to restore_state themselves, instead of simulation_loop doing it for them. That way users can do their own (re-)initialisations based on the restored variables. e.g. in malariasimulation, they may want to calculate an FOIM based on restored variables so that they can initialise their differential equation solver.

I'm sure those issues could be worked around by saving more state. My guess is that saving more state would be the most concise approach. But it's something to bear in mind.

Thanks so much!

[plietar]

Thanks for the review, I've applied the suggestions.

I think it would be much simpler to have malariasimulation save and restore its state buffers than re-creating them from the Variables (I'm not sure that would even be possible). It should be able to do that before/after calling simulation_loop, and let simulation_loop call restore_state to save the events and variables.

Another option would be to add an argument to simulation_loop for custom checkpoint-able objects, and have restore_state call checkpoint/restore on these automatically. Then the caller only has to define a pair of methods on each object but not worry about calling them.

[giovannic]

To me, the former option seems like a good place to start. Looking forward to it!