The difference between nrn_deliver_events and deliver_net_events

[Kim1230]

Hello,
I have a question to understand the operation of the coreneuron. First, can you explain what is the net_cvode_instance. And why for each fixed time step, it will have a deliver_net_events at the begining and have a nrn_deliver_events at the end ? Both of them will operate the _net_receive_kernel to update state, what is the difference? And I think delievr_net_events have finished to operate all the net_receive_buffer. What is the new net_receive_buffer for nrn_deliver_events? Hope for your help. Thanks!

[nrnhines]

Near the beginning of a time step, deliver_net_events checks thresholds and then delivers all events up to tentry+0.5*dt. This allows a sequence of 0 delay events consequential to a threshold crossing to be delivered immediately.

Near the end of time step, nrn_deliver_events delivers all events up to the end of the time step (no thresholds are checked).

net_cvode_instance is the unique singleton instance of the NetCvode class which is a wrapper for all the apparatus that manages a network simulation which involves time step integration and event delivery. In CoreNEURON, only fixed step integration
is supported but in NEURON this includes at the CVode variable time step variable order integration method. With CVode it is necessary to integrate just up to the earliest event, deliver that event, reinitialize the CVode integrator, and continue the integration to tstop or the next event.

[Kim1230]

@nrnhines Thanks a lot for helping me. now I wanna see output spikes associated to each cell. I can see generated spikes are associated to spikevec_time and spikevec_gid. I guess the gid is related to cell_id(for example, the network has 7742 cells so the gid will be 0-7741) while time is related to timesteps(0-15000). While the code of outputspikes sorts the spikevec_time and spikevec_gid. I wander the result of this sort. Is the relation of spikevec_time and spikevec_gid not changing ? What is meaning of this sort ? (I see it may be related to the MPI rank). Hope for your help!

[Kim1230]

And I also have a problem for the testing. When I use the same model, I set up mpi n=80 and mpi n=40. The result of number of spikes is different. It leads to the difference of population average rate. Can you tell me why ? Thanks a again

[nrnhines]

In the end, one want a spike raster file which is sorted by spike output time along with secondary sorting (in the case where multiple gids fire at the same time) by gid. It is up to the model author whether to 1) write each mpi rank spikes in separate files at the end of the sim, then concatenate and sort afterward. 2) at the end of the sim, write to a single file, then sort afterward. 3) use an elaborate parallel algorithm at the end of the sim to write a sorted spike raster file. Sorting an unsorted spike raster file after the sim can be done with

/where/you/installed/neuron/bin/sortspike unsorted_file_name sorted_file_name

where sortspike is

#!/usr/bin/env bash
sort -k 1n,1n -k 2n,2n $1 > $2

Since I don't know anything about your model, I can't say why the number of generated spikes is different.

You are correct in surmising that the results for 80 and 40 mpi ranks should be identical (after sorting the spike raster files).
If they are not, then the reason(s) for the bug(s) must be found. All debugging methods begin with the question, which is the first cell gid which has a different spike time in the two raster files?

[Kim1230]

@nrnhines Thanks a lot. I will try to find the problem of model. And what is the meaning of the spikevec_gid. Is that the cell_id for all cells generated by a model ?

[nrnhines]

whenever a spike occurs, one generally wishes to record both the spike time and the cell (gid) that spiked. A cell gid should be globally unique (two cells on different MPI ranks should not have the same gid)