Multi-core parallel support for JuMP supported solvers?

Specific Domains Optimization (Mathematical)
, , ,
16

@charshaw I modified the script a bit to run mwe twice; here are my results of @time optimize! within:

for n in {1..8} do                
OMP_NUM_THREADS=$n julia using_JuMP_multithreading.jl;

I fixed n = 450;

passing

scs_opt = optimizer_with_attributes(
    SCS.Optimizer,
    "linear_solver"=>SCS.MKLDirectSolver,
    "eps_abs"=>1e-8, # to increase the iteration count
    "verbose"=>false,
)

to mwe_program I get

Creating the problem instance... n = 450, OMP_NUM_THREADS = 1
  7.061236 seconds (1.04 k allocations: 90.613 MiB)

Creating the problem instance... n = 450, OMP_NUM_THREADS = 2
  5.475805 seconds (1.04 k allocations: 90.613 MiB)

Creating the problem instance... n = 450, OMP_NUM_THREADS = 3
  5.021955 seconds (1.04 k allocations: 90.613 MiB)

Creating the problem instance... n = 450, OMP_NUM_THREADS = 4
  4.565751 seconds (1.04 k allocations: 90.613 MiB)

Creating the problem instance... n = 450, OMP_NUM_THREADS = 5
  4.389557 seconds (1.04 k allocations: 90.613 MiB)

Creating the problem instance... n = 450, OMP_NUM_THREADS = 6
  4.413405 seconds (1.04 k allocations: 90.613 MiB)

Creating the problem instance... n = 450, OMP_NUM_THREADS = 7
  4.360352 seconds (1.04 k allocations: 90.613 MiB)

Creating the problem instance... n = 450, OMP_NUM_THREADS = 8
  4.339999 seconds (1.04 k allocations: 90.613 MiB)

so about 1.6 faster leveling around 4-5 threads. This comes from the following two lines: Code search results · GitHub

So either

  • you have plenty of psd constraints and project in parallel (mwe has only one), or
  • the (sparse CSC) matrix A defining the problem has enough density so that parallelizing A'x over its columns is beneficial.

Note: I don’t observe the same scaling with SCS.DirectSolver, solve time stays around 7s.

1 Like