The effect of execution of neuron when it has the coreneuron

[Kim1230]

Hello,
I have a problem for running the neuron with coreneuron code. When I debug it , I find the main process will execute Neuron's nrn_finitialize with neuron's mod file. But when it executes to the coreneuron's run_solve_core. It seems that all varaibles generated by the neuron will be regenerated by the coreneuron. I think this will consume much time and memory . If my understanding is wrong, please let me know. Thanks!

[nrnhines]

The round trip time cost should be on the order of the time it takes for NEURON to compute a time step. Whether that is presently 5 or 10 or 40 time steps in duration, I don't know. Since there are 40 (dt = 0.025) steps per ms, if tstop is much greater than a ms the round trip time is worth accepting if CoreNEURON can compute a time step sufficiently faster than NEURON and there are enough time steps between the round trip. In terms of memory, CoreNEURON memory use for a model is generally about 1/6 the size as for
a NEURON model. Transient memory during the round-trip copying is generally much smaller than that.

Now that CoreNEURON has been validated to generate same results for network models, we have begun the process of merging
the high performance benefits back into NEURON so that NEURON's data representation for the model adopts CoreNEURON's representation. When this is complete, there will be only one copy of the model data on the computer and that will be CoreNEURON's high performance version of the data. The repository master is tagged 9.0.dev and is the version for which a number of pull requests are under review that are moving this forward. In particular #1929 #2027

[Kim1230]

Hello@nrnhines, thanks for your response. I have a question and hope your response. I find that the network establishment and simulation of the contribution of ion channel or sypase mechanism according to xxx.mod will be executed in the neuron platform in the method nrn_finitialize() and setup_tree_matrix(). But in the coreneuron platform which will rebuild the network and rerun the simulation, direct_mode_initialize() will first clear_event_queue() and then nrn2core_tqueue().
The function clear_event_queue() I thought that it will clean all queues generated by the neuron simulation. But the function nrn2core_tqueue() copy each thread's queue from neuron to coreneuron. So, the thread's queue is not the event's queue? And the ion channel and sypase mechanism and setup_tree_matrix are needed to run by neuron and coreneuron seperately?
Hope for your response. Thanks!

[nrnhines]

In direct memory mode, CoreNEURON does not do the equivalent of nrn_finitialize but accepts the data from NEURON as already being initialized. That includes the event queues. This means you can run on the NEURON side (starting from finitialize) for a period of time (so that the NEURON queue may have many events that have not been delivered) and then start CoreNEURON at that final state to continue the simulation. Furthermore, when CoreNEURON finishes, then all data (including queues) are copied back to NEURON so NEURON can continue the simulation from there. The clear_event_queue on the CoreNEURON side that you mention does not clear the queue on the NEURON side. Conversely when CoreNEURON sends its data back to NEURON, NEURON will first clear its own queues and then copy the CoreNEURON queues. After saying this, it seems reasonable to clear the NEURON queue after it has been transferred in order to save space. I will look to see if this is already the case.

the thread's queue is not the event's queue?

Every thread in NEURON has its own (event)queue. Every thread in CoreNEURON has its own (event)queue.

setup_tree_matrix are needed to run by neuron and coreneuron seperately

If NEURON is simulating, then every time step NEURON is calling its setup_tree_matrix and CoreNEURON is idle, If CoreNEURON is simulating then every time step CoreNEURON is calling its setup_tree_matrix and NEURON is mostly idle (NEURON may be receiving a small amount of data from CoreNEURON on a per time step basis for GUI plotting.)