PipsNLP.jl is a Julia interface to the PIPS-NLP nonlinear optimization solver. Running the solver requires a working PIPS-NLP installation following the instructions.
At some point, PIPS-NLP updated which broke support for linking constraints. If you wish to model with linking constraints, try checking out the following commit hash from the
PIPS-NLP repo and compiling: 62f664237447c7ce05a62552952c86003d90e68f
Also note that you will need to compile PIPS-NLP with the same version of MPI used for your Julia environment.
PipsNLP.jl can be installed using the following Julia Pkg command.
using Pkg
Pkg.add(PackageSpec(url="https://github.com/zavalab/PipsNLP.jl.git"))or simply:
pkg> add https://github.com/zavalab/PipsNLP.jl.gitThe PipsNLP.jl package works with Plasmo.jl to model and solve optimization problems in parallel.
Currently, PipsNLP.jl supports three main modes for modeling. These are:
-
- Model in parallel and execute with mpirun
-
- Model in parallel using Julia's mpimanager where mpiranks correspond to Julia workers
-
- Model in serial and then distribute to julia workers and solve with MPI.
The examples folder has examples for each of these modeling approaches. While the first approach is probably the most familiar to MPI users, the second two approaches make it possible to interact with the PipsNLP model (e.g. to perform multiple solves in a Julia session).
The following snippet shows how one might model using this third approach to distribute an optigraph and execute PIPS-NLP.
using MPIClusterManagers # to import MPIManager
using Distributed # need to also import Distributed to use addprocs()
#setup worker environments
@everywhere using Pkg
@everywhere Pkg.activate((@__DIR__)) #change this to hit the correct environment
@everywhere using Plasmo
@everywhere using PipsNLP
#Setup MPI manager
manager=MPIManager(np=2) # specify, number of mpi workers, launch cmd, etc.
addprocs(manager) # start mpi workers and add them as julia workers too.
#create an optigraph with Plasmo.jl
graph = OptiGraph()
#Add optinodes
graph = OptiGraph()
n1 = @optinode(graph)
n2 = @optinode(graph)
@variable(n1, 0 <= x <= 2)
@variable(n1, 0 <= y <= 3)
@variable(n1, z >= 0)
@constraint(n1, x+y+z >= 4)
@objective(n1, Min, y)
@variable(n2,x)
@NLconstraint(n2,ref,exp(x) >= 2)
@variable(n2,z >= 0)
@constraint(n2,z + x >= 4)
@objective(n2,Min,x)
#link constraint between nodes
@linkconstraint(graph,n1[:x] == n2[:x])
#map julia workers to MPI ranks
julia_workers = sort(collect(values(manager.mpi2j)))
#distribute the graph to workers. This creates the variable `pipsgraph` on each worker with an allocation of optinodes.
remote_references = PipsNLP.distribute_optigraph(graph,julia_workers,remote_name = :pipsgraph)
#solve with PIPS-NLP and MPI
@mpi_do manager begin
using MPI
PipsNLP.pipsnlp_solve(pipsgraph)
end
#fill the local solution with the result on worker 2 (worker 2 is the root MPI rank)
PipsNLP.fill_solution!(graph, :pipsgraph, 2)