sPyNNaker
sPyNNaker copied to clipboard
Allow variation of timescale factor in run
It should be possible to set the time scale factor when run is called (this would be a SpiNNaker extension). This would then allow people to try a network, then reset it, then try again with a different timescale factor without having to re-map and load the network. Additionally, it would allow the user to run a number of timesteps with one timescale factor, then run the next set of timesteps with a different timescale factor. This would allow the system to cope with initial bursts of activity without becoming behind in timesteps.
Note this is relatively easy to implement with the current code; we just need to update the message that sends the run time to also include the machine time step.
We need to be careful here for a few reasons.
-
It is possible to do multiple runs between resets what happens here? Especial if second run call has no such factor?
-
Are we sure this does not change partitioning? Escpecial with the Sdram costs?
To answer the questions:
- I think the timescale factor should be the last value set in all these cases, and if not set initially, do what we do now (which is to read it from the config file, or set it automatically based on the time step). Note it should not be possible to change the time step itself, as this has a knock-on effect on everything. The time scale factor is just the speed at which the simulation is run.
- Because of the above, the time scale factor does not change partitioning or the SDRAM cost.
Probably now best left for a future release rather than this one.
@rowleya 2.because of the above, the time scale factor does not change partitioning or the SDRAM cost. not true when you consider roberts ear. It adjusts its partitioning based on the time scale factor. as it changes its ihcan n neurons to reflect how many it can support for the time fame allowed.
if you want to go down this road. it will need to be tied into a param settable in each pop and allow the vertex to decide if it needs remapping accordingly.