AEP: Distributed gap junctions

arbor-sim / arbor

The Arbor multi-compartment neural network simulation library.

https://arbor-sim.org

BSD 3-Clause "New" or "Revised" License

108 stars 60 forks source link

AEP: Distributed gap junctions #1766

Open noraabiakar opened 3 years ago

noraabiakar commented 3 years ago

Cells connected via gap-junctions currently have to reside on the same-cell group. As a result these cells cannot be distributed across multiple nodes. This is a major constraint for scaling up network sizes.

There are multiple approaches to distributing cell-groups including but not limited to:

Using waveform relaxation for solving the cell-group voltage matrices. This involves multiple rounds of communicating voltage values within a single epoch until convergence is reached among the cell-groups.
Communicating the voltages at every time step.

thorstenhater commented 3 years ago

A note to clarify the difference: an epoch is many timesteps in our terminology. Consequently WR requires some iterations of the same epoch, after each exchanging a trace (one entry per time step) of GJ voltages. The second approach requires exchanging voltages once per step.

max9901 commented 2 years ago

https://github.com/max9901/arbor/tree/exploration/distrgj

i did a exploration for this point:

Communicating the voltages at every time step.

The logic is their right now but the coding is definitely not up to spec, nor focused on any kind of performance. , next step is a proper rewrite and afterwards do some bench marking.

I would love some feedback on the idea, But i am aware there are a lot of memory inefficiencies and general very poor code in this exploration so please leave those remarks for later :D