[NEST Users] Re: Fastest way to transfer dense connectivity data in a parallel simulation?

Hi Robin, We are currently working with the Allen Institute to develop an efficient reader for large Sonata network specifications. We assume here that all connectivity is collected in HDF5 files, and expect significant performance gains if data is sorted by target neuron and the Sonata files provide "indices" tables. Would this help you? Dividing network data according to a specific compute node configuration seems rather restrictive to me and right now there is no way to read such data per process. Best, Hans Ekkehard -- Prof. Dr. Hans Ekkehard Plesser Head, Department of Data Science Faculty of Science and Technology Norwegian University of Life Sciences PO Box 5003, 1432 Aas, Norway Phone +47 6723 1560 Email hans.ekkehard.plesser@nmbu.no<mailto:hans.ekkehard.plesser@nmbu.no> Home http://arken.nmbu.no/~plesser From: Robin Gilbert De Schepper &lt;robingilbert.deschepper(a)unipv.it&gt; Reply to: NEST User Mailing List &lt;users(a)nest-simulator.org&gt; Date: Friday, 7 October 2022 at 15:41 To: NEST User Mailing List &lt;users(a)nest-simulator.org&gt; Subject: [NEST Users] Fastest way to transfer dense connectivity data in a parallel simulation? Hi all! In the world of biophysical detail, it's commonplace that the connectome is generated with algorithms that specify connections as dense tabular data, with each row specifying a synaptic location on a cell pair (SONATA for example). A) In NEST I can't really find the opportunity to fit this data into any of the connection rules: I want to specify pairwise connections from the multiset A to multiset B. - Is this possible with `pairwise bernoulli`, or do the inputs have to be strict sets? - The probability step is superfluous, can it be skipped? B) Then there's the fact that NEST parallelizes transparently, but since this data was generated in parallel by tiling the biological volume, I have neatly fragmented data already available on each node in the distributed cluster. It would be such a waste to communicate all the data to each node, for NEST to communicate and distribute them back another way. The data is too big to allgather and fit into memory of any single node. Not only is this a lot of overhead to implement, but NEST will throw away all but `1 / Nnodes` of the data on each node again, leaving me with a reshuffled version of my starting data. Is there a way to bypass the transparency and to imperatively declare the cells and connections on each machine? -- Robin De Schepper, MSc (they/them) Department of Brain and Behavioral Sciences Unit of Neurophysiology University of Pavia, Italy Via Forlanini 6, 27100 Pavia - Italy Tel: (+39) 038298-7607 http://www-5.unipv.it/dangelo/ Interested in large scale network modelling? Discover our framework<https://bsb.readthedocs.io/en/latest/>: [Image removed by sender.]<https://github.com/dbbs-lab/bsb>