Improve speed of simulations

[Tortar]

This gives a 20% speed-up in the single-threaded model at https://github.com/bancaditalia/BeforeIT.jl/blob/main/examples/benchmark_w_matlab.jl#L24, I expected the multi-threaded version to benefit more because we reduced also some allocations, but didn't test it

[Tortar]

Sorry I got fooled by some previous benchmarks, there is still some improvements but it is not 2x as I previously wrote, I edited my main post

[AldoGl]

Thanks @Tortar, I can see the trick is the "@view" macro.

Can I ask you what speedup do you get? On my machine I could measure around 6-10% for the standard Austrian parameterization (not bad!). Possibly it's even a bit higher for larger models

[Tortar]

same code as in #20

julia> @benchmark run(parameters, initial_conditions, T) # this PR
BenchmarkTools.Trial: 132 samples with 1 evaluation.
 Range (min … max):  36.550 ms …  41.297 ms  ┊ GC (min … max): 8.72% … 14.80%
 Time  (median):     37.836 ms               ┊ GC (median):    8.46%
 Time  (mean ± σ):   37.910 ms ± 815.271 μs  ┊ GC (mean ± σ):  8.76% ±  1.34%

      ▁    ▁▄▂ ▅▅▄▅▅▂█▂▁▄                                       
  ▅▁▃▁█▃█▆▆███▆██████████▅▅▆▅▃▅▁▃▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▃▁▁▁▁▁▁▁▃▃▃▁▁▃ ▃
  36.6 ms         Histogram: frequency by time         41.2 ms <

 Memory estimate: 110.18 MiB, allocs estimate: 670279.

julia> @benchmark run(parameters, initial_conditions, T) # master
BenchmarkTools.Trial: 118 samples with 1 evaluation.
 Range (min … max):  40.662 ms … 52.037 ms  ┊ GC (min … max): 8.02% … 6.88%
 Time  (median):     42.019 ms              ┊ GC (median):    7.65%
 Time  (mean ± σ):   42.583 ms ±  1.657 ms  ┊ GC (mean ± σ):  8.61% ± 2.31%

           █▂█ ▄ ▃ ▄                                           
  ▅▅▁▃▆▃█▆██████▆█▅█▆█▅▁▃▁▁▃▃▁▁▁▁▁▁▁▁▁▁▅▁▃▅▁▁▆▆▃▃▁▁▅▃▃▃▃▁▃▁▁▃ ▃
  40.7 ms         Histogram: frequency by time        46.6 ms <

 Memory estimate: 139.26 MiB, allocs estimate: 679000.

yes, nearly 10% :-)

[AldoGl]

Thanks @Tortar, squashing and merging now

[finmod]

@Tortar Can you see an advantage to bringing agents, agents_actions and markets of BeforeIT and ABCredit within JuliaDynamics/Agents.jl? This would provide better integration with the JuliaDynamics.jl-SciML,jl-ReinforcementLearning.jl ecosystems. Perhaps a ABM example in AgentsModelZoo...

See also this question on Discourse: https://discourse.julialang.org/t/specific-dom/121140

[Tortar]

Hi @finmod, I think that we should do the other way around if possible, try to base both of them on the Agents.jl framework. But I don't see this as a big priority and there are some challenges associated with this because of the origin and the current structure of BeforeIT.jl. I think if possible we should do that but this requires a lot of refactoring, and the current structure is not really that well supported from Agents.jl which uses an approach of arrays of structs while BeforeIT.jl uses struct of arrays (which could be more performant actually). Maybe we should allow the struct of arrays approach in Agents.jl. But this is a project which requires some efforts.

Personally, being one of the author of Agents.jl I would love to see this great models as part of the ecosystem, but for now it remains actually hard to accomplish. The advantages I see are that we could reuse some of the tools in Agents.jl and maybe more standardization in code style.

First step is understanding if we could support struct of arrays in Agents.jl, this adds complexity though on the framework side