HolyLow

@nrnhines No, the simulation fails with --multisend but without --gpu. The log is quiet similar: ``` $ mpiexec -np 8 ./x86_64/special-core --tstop 1000 --datpath ./networks/10000Sim/RoundRobin-core-8 --mpi --multisend libibverbs: Warning: couldn't...

@pramodk All the networks I use are generated by [Snudda](https://github.com/Hjorthmedh/Snudda/tree/master/snudda) and exported with the Neuron's nrnbbcore_write API. To meet your advice, I tested with a smaller network generated by Snudda...

@pramodk The original Snudda doesn't have the functionality of exporting network to CoreNeuron, and my modified version is too dirty to share. As a workaround, I uploaded my [generated tinySim...

@nrnhines I tried your datapoint, and without --multisend it worked fine on my environment as well. The outputs are also identical to what you've shown. But if I add the...

@nrnhines @pramodk Hello, I tried to dig into the source code, and the error seemed to come from the use_phase2_ procedure in nrnmultisend_setup.cpp. But I could hardly understand what the...

@pramodk Thanks for your attention. I was just trying to compare the behavior and performance when --multisend option was (or not) enabled. And I ran into the crash accidentally. When...

@nrnhines Thanks for your explanation! > At the time of dequeuing, the spike is place in a buffer specific to a mod file and at each time step that buffer...

@nrnhines I really appreciate your patience!! I have some more questions: > the cpu copies the spike to a modfile specific buffer on the cpu. Do you refer to the...

@nrnhines > That is very similar to the binqueue for NetCon events except td is in the binqueue so that spike.ts+netcon.delay is calculated only once. Wow, I haven't read the...

@nrnhines I've done some more detailed profiling, and it suggests that you are right about the GPU buffer allocations. The GPU buffers are not reallocated, so the performance is not...