Add timers for NEST 2.14.1

[jarsi]

This PR adds timer functionality to NEST 2.14.1. The usage and functionality is similar to what is implemented in PR #1778 and addresses the issues raised in #1471.

These timers are implemented:

• time_gather_spike_data
• time_simulate (basic timer, always enabled)
  • nested in time_simulate:
    • time_update
    • time_gather_spike_data
      • nested in time_gather_spike_data:
        • time_collocate_spike_data
        • time_communicate_spike_data
        • time_deliver_spike_data

In contrast to PR #1778, these timers are not implemented due to the infrastructure change:

• time_communicate_prepare (basic timer, always enabled)
  • nested in time_communicate_prepare:
    • time_gather_target_data
      • nested in time_gather_target_data:
        • time_communicate_target_data

I am not too familiar with the construction phases. Furthermore the order of collocate, communicate and delivery is different and I am not sure how this influences the time_gather_spike_data. So besides the normal review process, I'd ask specifically for input for these parts:

• time_construction_create (basic timer, always enabled)
• time_construction_connect (basic timer, always enabled)
• time_gather_spike_data

[jarsi]

Could one of you, @stinebuu or @hakonsbm, look at the use of time_construction_create and time_construction_connect? I could imagine these timers need some correction.

@jakobj, could you check whether time_gather_spike_data is implemented correctly?

[hakonsbm]

The time_construction_create timer looks fine.

For time_construction_connect there should be timers in the ConnectLayer function in topologymodule.cpp to get connection-times for spatial connections.

There should also be timers for DataConnect. Note that there are two high-level functions for DataConnect in nestmodule.cpp.

And see below for some other comments.

[jarsi]

Thanks for the review @hakonsbm. I added and changed the functionality as you suggested. Could you have a look again?

[hakonsbm]

Thanks! Looks good to me now.

[ackurth]

I ran some microcircuit benchmarks (10 s of model time) comparing the older custom version of the timers with the one from the PR for the time spent in the different phases during simulate.
Attached you find the data, all times are in seconds.
I looked at both installations with and without MPI support.
There seems to be a (reproducible - checked it multiple times) drop in performance using the new timers in both cases. Also the tripling in the communicate time might be noteworthy.
Otherwise, there seems to be a rather good agreement.

Benchmark	sim time	update time	collocate time	communicate time	deliver time
old timers, mpi=ON	7.99	3.13	0.24	0.30	3.34
old timers, mpi=OFF	8.45	3.87	0.63	0.03	3.54
new timers, mpi=ON	8.43	2.78	0.2	0.92	3.49
new timers, mpi=OFF	9.44	3.51	0.57	0.01	3.51

[jarsi]

Thanks for you measurements. The difference in communication time is due to a now missing synchronization point. Previously this synchronization was not measured and excluded from the communication time. Now it's implicitly contained in the call to communicate.

[jougs]

LGTM!

@ackurth: could you please indicate if @jarsi's last reply settles your concerns? Thanks!

[jougs]

Many thanks to everyone!

@terhorstd: let me know when you're in the mood for a release 🐱