search

tapios / risk-networks

Code for risk networks: a blend of compartmental models, graphs, data assimilation and semi-supervised learning
Other
2 stars 2 forks source link

`ContactSimulator` is too slow #61

Closed glwagner closed 4 years ago

glwagner commented 4 years ago

@lubo93 said:

Probably we have to go back to my original implementation where I simulated only one diurnal cycle (it's useful for system sizes N > 1e3). The current (continuous) contact simulation seems to be relatively slow.

We need to know the requirements of this system. What is the size of the network that we want to run, and how quickly do we need it to run?

lubo93 commented 4 years ago

Would be good to have something like 100,000 nodes for a "production" run at some point and for testing we need the 10,000 node network to estimate parameters. My old implementation needed something like 30 min for a 10,000 node simulation and a 90 day period (using approximated w_ji(t) evolution, i.e., not the actual master equation).

Would be good if we could also simulate about 90 days for a 10,000 node network in ~0.5 h.

glwagner commented 4 years ago

To start, we should maybe try just 30 days in 10 minutes of wall time, which will make it easier to run repeatedly. Is there a script that can be used for this purpose?