It's interesting that the link you included says " parallel barrier optimizer can not work deterministically because of design considerations". You might want to look at a different parameter, the parallel mode switch. The documentation (CPLEX 22.1) for that parameter says the opposite: "Parallel barrier optimization is only deterministic."
The issue seems to be with crossover rather than barrier itself. I don't know whether parallel threads make crossover nondeterministic, but I would not be surprised. You might try setting parallel to 1 (deterministic) and see if that matters.
Something to bear in mind when using multiple threads, particularly when you are bumping up against your CPU capacity, is that the operating system and background programs will be competing for those threads. If one system happens to have more stuff running in the background than the other does, that could make a difference. So I would look at how busy the respective systems are. Also, as I mentioned before, you should repeat the experiment at least once or twice to make sure this wasn't just random chance.
------------------------------
Paul Rubin
Professor Emeritus
Michigan State University
------------------------------
Original Message:
Sent: Fri May 05, 2023 01:12 AM
From: Sung Hwang
Subject: Different results with the barrier algorithm
All the parameter settings are the same on the both machines. Also, both machines have similar computing power.
The model is built based on live data, I can't repeat the run at this time.
Dr. Rubin, one quick question. As I explained in the above, I set mip.strategy.startalgorithm as 4 (Use barrier algorithm to solve the initial relaxation of a MIP) and use 16 threads for barrier algorithm.
Does using multi thread guarantee deterministic results?
According to the CPLEX 11.0 manual, if the threads parameter is greater than on, then opportunistic parallel MIP will be invoked.
http://www-eio.upc.edu/lceio/manuals/cplex-11/html/usrcplex/parallel5.html
Thank you for your help in advance.
------------------------------
Sung Hwang
Original Message:
Sent: Thu May 04, 2023 03:59 PM
From: Paul Rubin
Subject: Different results with the barrier algorithm
Just to be clear, all the parameter settings were the same on both machines? Also, did you repeat the runs and, if so, did A beat B again? Are the CPLEX versions the same on both machines? Does setting the random seed to the same value on both machines change the results?
------------------------------
Paul Rubin
Professor Emeritus
Michigan State University
Original Message:
Sent: Thu May 04, 2023 03:07 PM
From: Sung Hwang
Subject: Different results with the barrier algorithm
Hello,
I ran the same mip model on two different computers (Their specs are almost same) using the barrier algorithm. However, I got the totally different results.
Test on computer A, I got the solution in 847 sec, whereas the other computer B couldn't find the solution in 6800 sec.
I checked two log files and it seems that both computers took almost same amount of time on primal barrier. However the computer B spent much more time once dual cross over started.
I am little confused that both two computers are almost the same and solved the same mip models, but results are different.
Would anyone help me to understand this issue? Also, does barrier algorithm have stochastic properties that could lead to different results?
------------------------------
Sung Hwang
------------------------------