# Last edited on 2020-12-24 15:57:17 by jstolfi


NILTON'S BIG NETWORK

  Obtained a network description from Nilton Kamiji on 2020-12-03.
  The file is "data-nilton.txt" (9284524990 bytes, 9.3 GB).
  A description of it and of other network parameters is in
  Nilton's email "2020-12-03-nilton.msg".

  The network is a rough random model of a cortical column, after
  Potjans & Diesmann 2015. Conceptually the model has 4 layers, each
  with two populations of neurons (excitatory and inhibitory). The
  snapses are randomly generated so as tomee a specified number of
  connections btween each pair of popuations.

  This specific network has 77169 neurons in 8 groups and 298'880'968 synapses.

CONVERTING THE FORMAT

  The file was converted from Nilton's format to my {nmsim} format
  by the script {convert_nilton_to_nmsim.gawk} After many iterations
  (curst be he who invented languages without variable declarations)
  a file in correct format and apparently correct contents was 
  obtained: "data-nmsim.txt" (10205745917 bytes, 10.2 GB).
  
  The conversion discarded the delay parameter in each synapse, since
  the {nmsim}library does not support it.
  
SIMULATION

  The first successful simulation of the network used this command:
  
    time nmsim_elem_net_simulate \
      -netFile data-nmsim.txt \
      -timeStep 1.0 \
      -nSteps 1000 \
      -exInput 0 77168 0 1000  0.0 0.0 \
      -trace     0     0  350 850 \
      -trace 20683 20683  350 850 \
      -trace 26517 26517  350 850 \
      -trace 48432 48432  350 850 \
      -trace 53911 53911  350 850 \
      -trace 58761 58761  350 850 \
      -trace 59826 59826  350 850 \
      -trace 74221 74221  350 850 \
      -outPrefix out/2020-12-03

  Reading the input file took 300 seconds.  The simulation for 1000 time
  steps took 1250 seconds. Peak memory usage was 6.3 GB, just below the
  total free memory of my laptop after quitting Chromium. 
  
  Unfortunately the network was dead by time t=350, when 
  trace collection started.

  Made some improvements to the code to speed up reading and simulation.
  
MORE SIMULATIONS

  See the Makefile for the current simulation run command and options.

  An attempt was made to "fake" the refractory period by resetting the
  recharge modulator M[t] to M_R = 0 after a spike, recovering to 1
  with \tau_M = 2 ms.  Did not seem to help much.