Hello, we are Julia beginners. We want to use Julia for solving linear programs using JuMP and some external solver. For very large LP problems the model generation time can be significant.
Since JuMP covers many different options to store variables, constraints and coefficients of a model we found it important to know the speed and memory usage of the different options.
If you are experienced with Julia/JuMP, the optimization Packages (CPLEX, Gurobi, Mosek, Xpress) or with applied linear programming in general we would be very thankful if you could check the results and address the open questions(see below).
We mainly tried different container types:
Normal Julia arrays with one based indexing
DenseAxisArrayswhich allow string indexing
Normal Julia dictionaries
and also experimented with the direct_mode option.
The containers were applied to:
the JuMP variables ,
We used Julia v1.5.1 and JuMP v0.21.3 and CPLEX.jl v0.6.6 The same Data is used for all tests. We built a simple linear example model with 1000 objective variables x[1:1000] and 1000 constrains and want to maximize the objective function sum(x).
You can find the code we used to create the benchmarks at: “https://github.com/SonntagK/JuMPContainerBenchmarks”
There you can also find the results in the folder called Results as .xlsx files:
Here is a part of the results. The memory estimate, median time and number of samples is obtained using the BenchmarkTools.jl. Everything we wrote is wrapped in functions and we used the second benchmarking results as recommended. In the end you can find an example how I create the data and build the model. In the code I used for the benchmark results everything is seperated into modules.
Surprisingly JuMP DenseAxisArrays were much faster than normal one based indexing Julia arrays. We are happy (since we like string indexing), but wonder if there is some problem in our use or the implementation of the normal array container.
From reading JuMP presentations, we expected that using the direct mode will speedup the model generation. But we could not observe this. Is this correct ??
Possibly using some containers (e.g. string based indexing with 20 random characters) may slow down the speed of the solvers. Therefore, the benchmark of the model only generation is compared to the benchmark of model generation + solving the optimization problem. We did not find any significant impact of the container on the solver speed.
Quesition1: Why are Arrays so slow compared to other container types, espescially compared to DenseAxisArrays?
Question2: The direct_mode option does not help. Why? Are our models to small to see the impact?
For any further remarks or recommendations We would be very thankful.
Example for data creation, model build and optimization for denseAxisArray:
using JuMP, CPLEX # create random data n = 1000 arrayCoeff = rand(n,n) arrayBound = rand(n) #transform Data to denseAxisArray coeff = JuMP.containers.denseAxisArray(arrayCoeff,1:n,1:n) bound = JuMP.container.desnseAxisArray(arrayBound,1:n) #model build m = Model() set_optimizer(m,CPLEX.Optimizer) @variable(m, 0<=x[1:n]<=1, container = DenseAxisArray) @constraint(m, con[i = 1:n], sum(coeff[i,j]*x[j] for j in 1:n) <= bound[i], container = DenseAxisArray) @objective(m,Max,sum(x)) #optimization JuMP.optimize!(m)
edit1: Added link to results, added jpg of results and explanaition how the numbers arise
edit2: Added an example