search

lanl-ansi / rosetta-opf

AC-OPF Implementations in Various NLP Modeling Frameworks
Other
52 stars 11 forks source link

Add CasADi via PythonCall #25

Closed odow closed 2 years ago

odow commented 2 years ago

Closes #21

Alternative to #24

odow commented 2 years ago

I think this answers our questions.

Python call has faster callbacks (106 seconds in Ipopt instead of 180), but it also has a much lower overhead of ~50 seconds compared with ~145 for the NL file. Some of the overhead is writing and reading the NL file and JuMP, but it isn't an additional 100 seconds, or we would notice with JuMP vs NL files using Ipopt_jll.

casadi.jl

Total seconds in IPOPT                               = 106.723

EXIT: Optimal Solution Found.
       model  :   t_proc      (avg)   t_wall      (avg)    n_eval
       nlp_f  |  22.99ms (270.48us)   3.17ms ( 37.25us)        85
       nlp_g  |   3.40 s ( 39.98ms) 480.38ms (  5.65ms)        85
  nlp_grad_f  | 266.91ms (  3.07ms)  36.94ms (424.61us)        87
  nlp_hess_l  |  10.56 s (125.74ms)   1.49 s ( 17.77ms)        84
   nlp_jac_g  |   8.89 s (103.37ms)   1.25 s ( 14.56ms)        86
       total  | 760.02 s (760.02 s) 106.73 s (106.73 s)         1

Summary
   case........: /Users/oscar/Documents/lanl-ansi/rosetta-opf/variants/../data/pglib_opf_case10000_goc.m
   variables...: 76804
   constraints.: 112352
   feasible....: true
   cost........: 1354031
   total time..: 153.1173529624939
     data time.: 9.699323892593384
     build time: 35.126672983169556
     solve time: 108.29135513305664

jump-casadi.jl

Total seconds in IPOPT                               = 181.199

EXIT: Optimal Solution Found.

Summary
   case........: /Users/oscar/Documents/lanl-ansi/rosetta-opf/variants/../data/pglib_opf_case10000_goc.m
   variables...: 76804
   constraints.: 112352
   feasible....: true
   cost........: 1354031
   total time..: 325.1633548736572
     data time.: 9.590620040893555
     build time: 8.191709041595459
     solve time: 307.3810248374939