Originally posted by: danielBF

Is it possible to formulate a model in Java using parallel threads?
Could that improve the performance of building the model?
Or is that impossible due to the used interface?

One example could be the generation of restrictions in seperated threads.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: RWunderling

Unfortunately, this is not supported by any of our interfaces. CPLEX is thread safe in the sense that you can build different models with different CPLEX / CPXENV+CPXLP (dependent on the API you are using) object, but you cannot concurrently operate on the same object.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: danielBF

Disagreeable to hear that.
Many thanks for your quick response.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

With OPL, you can use multiple threads and have 1 OPL instance by thread. To get this, you must ensure you have 1 OPL factory by thread. (see the portfolio example in OPL sameples suite).
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: SystemAdmin

Are you still looking into this? If so, then I suggest to take a look into the CplexServer.java example shipped with CPLEX. This shows how you can create a model and transfer it over an ObjectStream. Similarly, you could create a model and transfer it into an instance of IloCplex. And I think this will work in parallel to some degree.
The following program builds up two disjoint parts of the model in two different threads (I just used two different IloModeler instances and did not bother to actually create the threads) and then collects the per-thread models into one instance of IloCplex:

import ilog.cplex.*;
import ilog.concert.*;
 
public final class ParMod {
   public static void main(String[] args) {
      try {
         // First "thread".
         IloCplexModeler m1 = new IloCplexModeler();
         IloNumVar[] x1 = new IloNumVar[2];
         x1[0] = m1.numVar(0, 1, "x1");
         x1[1] = m1.numVar(0, 1, "x2");
         IloRange[] rng1 = new IloRange[1];
         rng1[0] = m1.ge(m1.sum(x1[0], x1[1]), 1, "c1"); // x1 + x2 >= 1
 
         // Second "thread".
         IloCplexModeler m2 = new IloCplexModeler();
         IloNumVar[] x2 = new IloNumVar[2];
         x2[0] = m1.numVar(0, 1, "x3");
         x2[1] = m1.numVar(0, 1, "x4");
         IloRange[] rng2 = new IloRange[1];
         rng2[0] = m2.ge(m2.sum(x2[0], x2[1]), 2, "c2"); // x3 + x4 >= 2
 
         // Now collect everything into one IloCplex instance.
         IloCplex cplex = new IloCplex();
         cplex.add(x1);
         cplex.add(x2);
         cplex.add(rng1);
         cplex.add(rng2);
 
         // Add constraint and objective that have variables from both
         // sub-models.
         IloRange c3 = cplex.addEq(cplex.sum(cplex.sum(cplex.prod(1, x1[0]), cplex.prod(2, x1[1])),
                                             cplex.sum(cplex.prod(3, x2[0]), cplex.prod(4, x2[1]))),
                                   8, "c3");
         cplex.addMinimize(cplex.sum(cplex.sum(x1[0], x1[1]),
                                     cplex.sum(x2[0], x2[1])));
 
         // Export model to file and solve it.
         cplex.exportModel("parallel.lp");
         cplex.solve();
 
         // Print results.
         System.out.println("x1[0] = " + cplex.getValue(x1[0]));
         System.out.println("x1[1] = " + cplex.getValue(x1[1]));
         System.out.println("x2[0] = " + cplex.getValue(x2[0]));
         System.out.println("x2[1] = " + cplex.getValue(x2[1]));
         System.out.println("slack[c1] = " + cplex.getSlack(rng1[0]));
         System.out.println("slack[c2] = " + cplex.getSlack(rng2[0]));
         System.out.println("slack[c3] = " + cplex.getSlack(rng1[0]));
         System.out.println("objective = " + cplex.getObjValue());
 
      } catch (IloException e) {
         System.err.println(e.getMessage());
         e.printStackTrace();
         System.exit(-1);
      }
   }
}

Note that the stuff that is build up by the different threads is disjoint. I don't know how far you can carry this but it might be worth a try.

Depending on what costs the most time in building up your model you could also try the following: Create all variables that exist in the model. Now, instead of building up the constraints as instances of IloNumExpr build up each constraint as an array of IloNumVar and an array of double (I don't discuss lower and upper bounds for constraints since they are handled similarly). For example, an expression like x1+2x2+3x3 would be stored as

new IloNumVar[]{ x1,  x2,  x3 };
new double[]   { 1.0, 2.0, 3.0 };

This can be done in parallel. When you have build up all constraints like this then do one final sequential pass over all these arrays and use IloCplex.scalProd() to create instances of IloLinearNumExpr for each constraint. Depending on how much time you save in the parallel setup of the arrays this may or may not be faster than setting up everything sequentially.
#CPLEXOptimizers
#DecisionOptimization

Originally posted by: Mahesh Bakshi

Does this mean, it is not possible to solve in parallel cplex models written in java as shown in this link http://stackoverflow.com/questions/36094529/unable-to-use-cplex-java-class-in-parallel , 

(sorry for using external link to ask question)

Thank you in advance.

#CPLEXOptimizers
#DecisionOptimization

Hm, the error reported in this stackoverflow question is completely unrelated to multi-threading. The error says that the CPLEX library could not be found. That is a configuration problem that does not depend on the number of threads you use. It looks like the code in SO uses MPI. So maybe it involves some hosts on which CPLEX is installed in a different place or not at all?

Anyway, it is definitely possible you have multiple solves in parallel. You just need a separate IloCplex instance for each solve, then you should be good.

#CPLEXOptimizers
#DecisionOptimization

Originally posted by: Mahesh Bakshi

Thank you Daniel

#CPLEXOptimizers
#DecisionOptimization