Create a toolbox to check performance

[sandcha]

Find common ways to analyse OpenFisca:

• Time performances
• Memory performances
• Calculation cache

• Define common metrics to check performances evolutions in Python and Web APIs

• Stack depth?
• Cache hits and misses? (Cached but unused values, ...)
• Percentage of default values in calculation
• Number of calls to calculate()
• How many times a calculate() returns 0 or any equivalent value to false (% unnecessary computations in conditions)
• Inputs impact in calculation performance (number of calculations to a variable listed in the inputs but at a different period from the calculation period)

• Unify Core tracing process with:

• computation_log()
• Tracer stack
• simulation stack for cycle and spiral detection

➡️ See #884

• Implement computation log as tree structure with each node registering

• Done:
  • variable name
  • period
  • wall-clock time
  • computed value
  • (children node)
• Still to be done:
  • value originating in: (input, calc, cache, default, extinct formula)

• Post-process the tree to bring out derived metrics:

• stack depth at each node
• time spent on each node
• number of children computations of each node
• percent of children node coming from each of cache, input etc.

• Check metrics possible representations

• Generate a graph similar to profilers (icicle ?) with variables and parameters?

➡️ See #895

[bonjourmauko]

Are we still looking forward to make this happen? Recently, another initiative has emerged based on inputs and contributions by @benjello @sandcha @benoit-cty and @clallemand. We could maybe merge/supercede #1315