Test with ExaModels

[jbcaillau]

@0Yassine0 @ocots check #25

[jbcaillau]

Naïve try with julia -t 4 on an Inria cluster¹.

@amontoison if you can run the code on aurora please (increase Nand) proceed 🥲

julia> include("goddard-exa.jl")
NLP + Ipopt :  91.587 ms (428357 allocations: 22.29 MiB)
NLP + MadNLP:  98.770 ms (479052 allocations: 69.30 MiB)
Exa + Ipopt :  44.883 ms (792 allocations: 81.62 KiB)
Exa + MadNLP:  57.040 ms (5516 allocations: 50.80 MiB)

NLP + Ipopt : Generic Execution stats
  status: first-order stationary
  objective value: -1.0125714090309206
  primal feasibility: 8.783421906466416e-10
  dual feasibility: 1.5628074497679174e-9
  solution: [1.0  2.8447284383565784e-40  1.0  0.9999813724318203 ⋯ 0.20197397954338486]
  multipliers: [3.942852127522442  0.14628571080470762  0.0541231604715201  3.9406167167698016 ⋯ -0.08152196877705907]
  multipliers_L: [0.25059035470884994  0.25059035596800616  0.0  2.5059502143790325e-9 ⋯ 1.240706076116897e-8]
  multipliers_U: [0.0  2.505903309089731e-8  0.0  0.00013445442754716537 ⋯ 0.0]
  iterations: 18
  elapsed time: 0.092
  solver specific:
    real_time: 0.09275698661804199
    internal_msg: :Solve_Succeeded

NLP + MadNLP: "Execution stats: Optimal Solution Found (tol = 1.0e-08)."
Exa + Ipopt : Generic Execution stats
  status: first-order stationary
  objective value: -1.012571409018816
  primal feasibility: 1.8874651026964864e-10
  dual feasibility: 1.1131274941257247e-10
  solution: [0.20197397128380745  1.0  1.2694438541840876e-33  1.0 ⋯ 4.39107280547744e-6]
  multipliers: [-1.240703620864565e-8  4.193737933722494  0.40085780685347294  0.05360611481410709 ⋯ 0.08152196564186191]
  iterations: 18
  elapsed time: 0.049
  solver specific:
    real_time: 0.04976511001586914
    internal_msg: :Solve_Succeeded

Exa + MadNLP: "Execution stats: Optimal Solution Found (tol = 1.0e-08)."

Footnotes

1. 40 CPUs (10 cores each), info below:
   model name : Intel(R) Xeon(R) Silver 4114 CPU @ 2.20GHz
   stepping : 4
   microcode : 0x2006b06
   cpu MHz : 2200.000
   cache size : 14080 KB
   physical id : 0
   siblings : 20
   core id : 2
   cpu cores : 10 ↩

[amontoison]

@jbcaillau ExaModels.jl was developed for GPUs, I don't think that you will see a difference if you use miltiple threads.
Can you also test with one thread?

[jbcaillau]

@amontoison I did, no change. again, at this step this is completely naive. just tried to follow hints below (from ExaModels doc):

the point with ExaModels is to use (i) multi-threaded CPUs and / or (ii) GPUs as documented here.

if you can trivially run the code to test speed up on GPU please let us know 🤞🏽

[amontoison]

I think that you need backend=CUDABackend() to easily test it on NVIDIA GPUs.

[jbcaillau]

I think that you need backend=CUDABackend() to easily test it on NVIDIA GPUs.

@amontoison yes. just tested it on our Inria cluster. (need more time to eval performance.) check #25 (comment)

[jbcaillau]

New run: #25 (comment)