SuperSpikeConnection in parallel

Discussions on development, Auryn's inner workings, adding new features and how to change things at the core
Post Reply
rob_luke
Posts: 1
Joined: Thu Oct 18, 2018 5:45 am

SuperSpikeConnection in parallel

Post by rob_luke »

Dear Friedemann,

First, congratulations on the new position.

I am just wondering if you have plans to support parallel processing for the superspike repo? I understand with moving labs etc that this might not be a high priority at the moment.

I find I can use up to 8 CPUs, but above that evolve_locally is false and things don't get initialised.
User avatar
zenke
Site Admin
Posts: 156
Joined: Tue Oct 14, 2014 11:34 am
Location: Basel, CH
Contact:

Re: SuperSpikeConnection in parallel

Post by zenke »

Thanks and thanks for reporting this. Time has indeed been a constraint lately. The answer to your question depends a bit on which connection class you are after. The random feedback connection should in principle support parallelism, so if you are seeing a problem there, this might be due to a bug. The original SuperSpike connection with symmetric feedback, however, should throw an exception when trying to run in parallel because it isn't fully supported there. This is due to fundamental design constraints. In Auryn the synaptic weights reside on the same process/node as their corresponding postsynaptic neuron. Computing this transpose as it is required to back-propagate through symmetric weights, requires all weights to reside on the same node/process. I have tried to hack around with this to remove this constraint, but I thought I had not made this code publicly available yet (since it wasn't working properly). Maybe you can elucidate a bit more what you are trying to run and where it crashes and I will see if I can reproduce your problem.
Post Reply