Originally posted by: geffer

Totally same code (column generation)  in java with cplex package 12.6

and I copy it from my small computer (windows system, 4 core, 8 G)to the super computer

(Unix 32 32G), to my surprise, the result is totally different between this two environment.

Is that common or I did something wrong during this process?

Does anybody help me to explain why?

Thanks

Keji

You need to clarify two things:

1. Are you trying to copy compiled binaries to the Unix box, or did you transfer source code and recompile there?
2. What is different about the "totally different" result? Is the problem feasible on one machine and infeasible/unbounded on the other? Do both produce optimal solutions, but with significantly different objective values? Are they both producing the same objective value but with different values for the decision variables?

Originally posted by: GUS_FC

Hi PaulRubin

I have similar question with more details. I have two different solution with different objective function values in Windows and Linux.

1. I copied compiled binaries to Linux box from Windows. 

2. Both environment provide feasible solutions. Both provide best solution found in terms of target gap (or optimality gap). The gap I used is 0.0005. All parameters used are the same in both environments. The values for the decision variables are different.

What would be the cause of this situation? I would appreciate any comment. 

1. Did you run your Windoze binaries with WINE on the Linux box, or run them in a virtual machine (with some version of Windows as the VM guest system), or what? I'm not sure it matters in terms of answering the question, but it might.

2. Apparently your problem has multiple optimal solutions. When that happens, it's fairly random which one you get. I'm not sure about LPs, but with MIPs some of the CPLEX internal logic uses pseudorandom numbers. I don't know whether the same seed would be used on both machines, and even if so I don't know whether the pseudorandom numbers would be identical (since I don't know how they're being generated). Also, under some circumstances really small differences in floating point arithmetic might change some choices. If you're running multithreaded (and maybe even if you're not), slight differences in timing could make a difference, especially with a MIP (where node A might complete before node B on one machine but after node B on another).

This is all pretty speculative on my part. I can't say I'm shocked, though. As I said, when you have multiple optima ("optima" defined to within tolerances), it's anybody's guess which one you get.