SPSS Statistics

Why Open Source is not the answer to everything

By Archive User posted Fri October 21, 2016 03:24 PM

  
Today we more and more hear about and tend to use Open Source software. In many cases this might be a viable and good choice. Apparently, mathematical optimization (aka math programming, aka decision optimization) usually is not one of those cases.

Here is a real client story. The client is a major car manufacturer in Germany. Some months ago, they asked one of their common project implementation suppliers to create a new optimization application for one of their business problems. This supplier then implemented an optimization approach that decomposes the original problem into thousands of optimization problems. They then tried an Open Source solver for this problem and this one seemed to do ok on the given problems, at least at first sight.

However, after doing some more testing, it turned out that for real problem sizes this approach had a severe problem: solving all the optimization models on a given hardware would have taken 65 -- days! Yes, indeed, more than 2 months. And this was obviously unacceptable for the business users.

The client then reached out to IBM and asked us whether they could try their models with CPLEX. And we gave them a CPLEX evaluation license and then they started testing with CPLEX. And then the first reaction was (literally): "We now tested the first models but this was so fast that we believe something went wrong".

But actually it was the contrary: nothing went wrong, CPLEX was able to solve the given models to the optimum (on the same hardware, of course) in very short time. The full set of models took less than 2 hours with CPLEX instead of 65 days with the Open Source solver before!

And guess what: line of business was so happy that they purchased CPLEX licenses right away :-)

Now, please: Check out IBM Decision Optimization on Cloud and see us at IBM World of Watson in Las Vegas in sessions 1675A, 2302A, 2619A, 3091A, 3587A, 3782A, and 3921A!

#DecisionOptimization
#MathematicalOptimization
#Optimization
4 comments
5 views

Permalink

Comments

Wed January 29, 2020 06:27 PM

Regarding

>>>>
So how did the openess of the source code of the "Open Source Solver" affect the performance so badly? Or the other way around, was is really the closeness of the CPLEX code that gave the performance boost?
>>>>>

I don't think the open or closed aspects explains the difference in performance.    The commercial version like CPLEX had more resources put into it and had a longer development period, so it was faster.    Furthermore, unlike something like the open source nature of the Linux operating system, the number of people with sufficient experiences to contribute to an open source LP/MIP solver is orders of magnitude smaller than the number of the CS knowledge to contribute something to an operating system.   Therefore, the resources devoted to an open source LP/MIP solver may be less than a closed, commercial one, while that won't be the case for an open source operating system.    So if you are considering a particular application and want to know whether an open source version "is the answer", ask yourself the question of how many people are capable of contributing effectively to an open source project.    The smaller that number, the less likely that the open source project will be the answer.

Wed October 26, 2016 09:56 PM

Stefan, in our business the open source solvers are the free available ones, while commercial solvers are closed-source (for various reasons). And as our experiences show, incl the example above, it is the commercial solvers that outperform the free ones.

Mon October 24, 2016 01:38 PM

So how did the openess of the source code of the "Open Source Solver" affect the performance so badly? Or the other way around, was is really the closeness of the CPLEX code that gave the performance boost?

Sat October 22, 2016 04:22 PM

Great and powerful story, thanks for sharing!!