Skip to content

Latest commit

 

History

History
 
 

NetworkOutput

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 

Back to Projects List

Storage of large-scale network simulation output

Key Investigators

  • Kael Dai (Allen Institute) @kaeldai
  • Ben Dichter (Stanford Lab) @bendichter
  • Yazan Billeh (Allen Institute) @CellAssembly

Project Description

An efficent, parallizable way to store the results, and even the input, of large-scale in-silico network simulations.

Objective

  1. Create an extension for storing and reading discrete single unit simulation data (ie spike times from a biophysical/point network simulation).
  2. Create an extension for continuous data (ie membrane potential, calcium conc).
    • May have to create a second extension for multi-compartment reporting (ie membrane potential along all sections of the dentrites)

Approach and Plan

Spike time recordings.

Goals: We want to be able to save spike trains from a simulation of anywhere between hundreds to millions of different cells. The previous way to do this was given every cell their own dataset.

old nwb spike format

While simple to understand, having to open a dataset handle for every possible cell in the network didn't scale well. For the new format, we store all spikes in a single file, using an index table link a given spike time with a cell gid.

new nwb spike format

Multicell, multicompartment continuous data storage

Goals:

  • Store cell variable data (membrane potential, [Ca++], etc), collected during a simulation
  • Multiple cells
  • Individual cells may be made up of different sections (soma, axon, dendritic branches) potentially needing to be stored.
  • Want to be able to write/read in parallel.
  • Want to be able to chunk data by time or by cell

Conceptural Framework:

  • Index table stores range of segment ids.
  • Stores relative recording position of each cell segment.

indexing multi and cell compartment cells

Segment recordings from each cell are stored in a single TimeSeries

accessing of data

actual example

Progress and Next Steps

  • Need to be able to use region references as a more explict way of indexing tables.

  • Critical: we need to be able to preallocate memory blocks and write in chunks (and to take advantage of MPI).

Illustrations

Background and References