Originally posted by: Job_Jonah

Hi,

I am working on a traditional benders decomposition. When I run it on my Windows computer in Cplex c++ API 12.6, it just works fine and master and subproblem converges at the optimal solution very quickly within 200 iterations.

However, when I run the exact same code at Linux operating system with  Cplex c++ API 12.6.2 (different version) at my university's high performance computing center, master and subproblem converge very slowly and reach to maximum iteration limit (5000) without reaching to the optimal solution. 

When I compile the deterministic version of the benders decomposition algorithm, it again works fine on linux operating system and gives me the optimal solution. 

What might cause the convergence problem in linux operating system? Does it happen due to having different version of CPLEX on Linux operating systems or parallel computing?

I really appreciate your help.

Best,

Jonah

This may have nothing to do with your problem, but how many threads are you running with on the Linux machine?  How many cores does that machine have?  What happens if you reduce the number of threads used on the Linux machine (e.g., to 12 or fewer, or the same as on the Windows machine)?

Originally posted by: Job_Jonah

The high performance computers (hipec) have 32 cores and windows computer has only 1 core. I reduced the number of threads to 1, 2, 4, 8, 16 while running it on hipec and tried to solve the problem again. The results are worse than when I have 32 threads. 

I also tried the following settings:  -1 core, 1 thread;  -1 core, 32 tread. It seems that it did not improve the result.

First of all, this is not unexpected behavior.

Using different versions of CPLEX on different machines can give (very) different running times. This can swing both ways and is known as performance variability.

To better understand the difference of the two runs I'd start by comparing the logs, potentially with additional output from callbacks. Then check things like

- does one run require significantly more nodes?
- does one run have a much worse node throughput?
- does one run find candidates for feasible solutions much more often?
- other things like that.

These things are especially interesting if you compare single-threaded runs. Maybe you can post of the two single-threaded runs here?

Originally posted by: Job_Jonah

Hi Daniel,

I have figured out the problem. Actually @rkersh 7a1ce5ff-8650-4ae4-a67f-69adbfe5e2f6 was right in his in his earlier response. The problem was the number of threads that are running on Linux machine. Other than master and sub problem, I have one more problem section that also feeds the master problem. What I realized is number of threads should not only be reduced for the master problem but also for the sub problem and other problem section. That solved my problem. Now they are giving me the same answer. 

Thank for the help.

Jonah