Remove set sender node id from deliver events pr

[suku248]

Retrieving the node id from SourceTable is computationally costly, and most targets don't need the node id (with the exception of binary_neurons). To support binary_neurons, an alternative solution is suggested here: The location (tid, syn_id, lcid) of the sender node id in SourceTable is stored in Event instead of the node id retrieved from SourceTable, and Event now provides a function to perform the look-up of the node id in SourceTable.

For local receivers (i.e., recording devices), delivery anyway takes a different, more direct path:

nest-simulator/nestkernel/event_delivery_manager_impl.h

Line 87 in 8ea4121

kernel().connection_manager.send_to_devices( tid, source_node_id, e );

The node id is retrieved from the local sender and stored in Event, and it can subsequently be retrieved through Event::get_sender_node_id().

[hakonsbm]

Looks pretty good to me, I have just a small suggestion.

[jakobj]

looks great, thanks for implementing this! 🚀
i have just one small comment, regarding the use of = for SpikeData, but maybe i was blind.

another thing i'd like to mention: as far as i understand, in binary neurons the gid is just used distinguish different presynaptic partners in the receiver node. wouldn't it be possible to instead use the spike data content to do that without looking up the gid? this would reduce (unpredictable?) memory access, but likely require more computation as the indices would need to be somehow combined into a unique identifier for each presynaptic neuron.

[jakobj]

sorry, i couldn't see where this was needed?

[suku248]

This was to silence a compiler warning through GitHub Actions. If I remember correctly, the problem was that since C++11 if a custom copy constructor or destructor is defined, a custom operator= also needs to be provided or it should be explicitly defined as default.

[jarsi]

Thank you @suku248 very much for this PR. I ran some benchmarks of this PR and the corresponding master on JURECA DC using a static version of the hpc benchmark. This PR is faster than master on all tested scales (1 to 64 nodes, 4 MPI processes per node, 32 threads, 250.000 neurons per Node). The performance difference is around 20%.

[suku248]

Thanks @jarsi for running some performance tests for this!
And thanks @hakonsbm and @jakobj for the review!

@jakobj the solution for binary neurons that you suggest sounds promising. Could you document it in an issue?

[hakonsbm]

Thanks!