In neurongpu.cu(Line 524 & 528) & spike_mpi.cu(320 & 351)，If use "MPI_Isend" & "MPI_Irecv"(Not blocking)，may be not recv data(lose data)?Is that right?

I would like to run the simulation under different conditions automatically with a loop. It looks like this :

for i in range(times):
net = Network(sim_dict, net_dict, stim_dict)
net.create()
net.connect()
net.simulate(sim_dict['t_sim'])
SaveData()

However, when a simulation is completed, the next one would be stuck at net.create() , as ValueError: Nodes cannot be created after calibration.
I think we need a reset operation like nest.ResetNetwork to initialize the network before we start to create it again, but I could not find it. Is there anything like this, please?

We are inconvenient to use sudo apt to install packages and would like to configure NeuronGPU under Anaconda3 python3.8 with TITAN XP, CUDA 11.2, then how could we configure the running environment, please?

Differently from NEST, in the current version of NeuronGPU nodes are not distributed automatically among MPI processes. The user must decide in advance how they are distributed, create nodes that must be connected remotely using the function RemoteCreate and connect nodes in different MPI process using the function RemoteConnect. Unfortunately this is a relatively new feature and is not yet documented.

like pynest，can NeuronGPU support this?（memory waste heavily）

I considered a simple 3 neurons system to mimic thalamus —> cortex and cortex <—> cortex interactions:

Axon delay : t_a = 1.5
Dendritic delay : t_d = 0.5
dt_stdp_update = t_post + t_d - (t_pre + t_a) = ∆t

Network:

"aeif" neurons

• TH : Thalamic neuron
• CX1 : Cortex neuron 1
• CX2 : Cortex neuron 2
• SG : 3 spike generators

Connections:

TH —> CX1 , CX2
CX1 <—> CX2
SG0 —> TH
SG1 —> CX1
SG2 —> CX2

Parameters:

W_th_cx_0 = 100.
Wmax_th_cx = 105.
lambda_th = 0.03

W_cx_cx_0 = 0.01
Wmax_cx_cx = 150.
lambda_cx = 0.1

Spike generators are set to make the network firing at times:

TH Spike Times = [100.5 , 200.5 , 300.5 , 400.5] ms
CX1 Spike Times = [102.5 , 202.5 , 302.5 , 402.5] ms
CX2 Spike Times = [104.5 , 204.5 , 304.5 , 404.5] ms

Since t_d - t_a = -1 ms every 100 ms (we neglect 100 ms ISI):

1. TH —> CX1 : ∆t = 2ms -1ms > 0 connection should POTENTIATE (OK)
2. TH —> CX2 : ∆t = 4ms -1ms > 0 connection should POTENTIATE (OK)
3. CX1 —> CX2 : ∆t = 2ms -1ms > 0 connection should POTENTIATE (connection get depressed)
4. CX2 —> CX1 : ∆t = -2ms -1ms < 0 connection should WEAKEN (updates correctly negatives ∆t but neglects potentiations (∆t=97. ms))

although thalamo - cortical connections are potentiated correctly, cortico-cortical connections are bidirectionally depressed making it impossible to form pairs of coupled neurons

I attach a zip with code and output of this system " issue_case_1.py
and the code to run the same model with only CX firing "issue_case_2.py"
Issue_3neurons_NGPU.zip

"iaf_psc_exp/iaf_psc_exp_g/iaf_psc_exp_hc" all these update dont use "weight"?But in nest，
V_.weighted_spikes_ex_ = B_.spikes_ex_.get_value( lag );
V_.weighted_spikes_in_ = B_.spikes_in_.get_value( lag );
S_.i_syn_ex_ += V_.weighted_spikes_ex_;
S_.i_syn_in_ += V_.weighted_spikes_in_;

if not use ”weight“，then send/recv spike will be meaningless.Am I wrong, or is it not implemented?