Profile subexperiment generation

[caleb-johnson]

With an increased focus on utility-scale experiments, we need to profile our subexperiment generation to ensure there are no easily-correctable inefficiencies. I have noticed that for cutting schemes involving lower-rotation gates, subexperiment generation can take a very long time when doing a full sampling, even for modestly-sized circuits.

[garrison]

I noticed that qiskit has benchmark tests based on asv in its test/benchmarks/ directory. We should add things like a benchmark related to this issue to a similar place in our repository, so that we can track performance over time, and at least have an idea of the key workflows where we care about performance.

[caleb-johnson]

Total subexperiments created: 12960
Total number of instructions in each subexperiment: Approximately 20-25

   344701052 function calls (303964246 primitive calls) in 149.617 seconds

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000  149.776  149.776 {built-in method builtins.exec}
        1    0.000    0.000  149.776  149.776 <string>:1(<module>)
        1    0.120    0.120  149.775  149.775 cutting_experiments.py:40(generate_cutting_experiments)
  12963/3    0.009    0.000  132.468   44.156 passmanager.py:417(wrapper)
        2    0.000    0.000  132.468   66.234 passmanager.py:127(run)
        2    0.000    0.000  132.468   66.234 passmanager.py:172(run)
        2    0.029    0.014  132.465   66.232 parallel.py:104(parallel_map)
    12960    0.228    0.000  132.434    0.010 passmanager.py:317(_run_workflow_in_new_process)
    12960    0.091    0.000  131.421    0.010 passmanager.py:270(_run_workflow)
25920/12960    0.148    0.000  119.654    0.009 base_tasks.py:202(execute)
   103680    0.490    0.000  119.071    0.001 base_tasks.py:72(execute)
    25920    0.461    0.000  116.043    0.004 dag_fixed_point.py:28(run)
31002016/373680   31.174    0.000   83.540    0.000 copy.py:128(deepcopy)
2382912/329184    6.889    0.000   81.959    0.000 copy.py:259(_reconstruct)
1516320/38880    4.007    0.000   80.666    0.002 copy.py:227(_deepcopy_dict)
5218560/2813120    6.502    0.000   59.745    0.000 copy.py:210(_deepcopy_tuple)
5218560/2813120    2.995    0.000   56.073    0.000 copy.py:211(<listcomp>)
928800/397440    1.667    0.000   56.064    0.000 copy.py:201(_deepcopy_list)

[garrison]

It appears that this is coming from the following lines, which were added in #458. Is that correct?

https://github.com/Qiskit-Extensions/circuit-knitting-toolbox/blob/21237168818ae89d0bb54002e260a06312d04b8f/circuit_knitting/cutting/cutting_experiments.py#L169-L188

[caleb-johnson]

It appears that this is coming from the following lines, which were added in #458. Is that correct?

https://github.com/Qiskit-Extensions/circuit-knitting-toolbox/blob/21237168818ae89d0bb54002e260a06312d04b8f/circuit_knitting/cutting/cutting_experiments.py#L169-L188

I am fairly certain. I am a little puzzled as to why dag_fixed_point is so expensive, and I don't see the other passes explicitly named here. Maybe the other run calls are those passes but with a generic name? I'm not sure right now tbh

This is obviously a truncated output. Maybe the other passes are named further down, but I didn't see them on quick glance