What is the efficiency of using CPLEX with C#.NET vs C++?

Question

Provided our UI must be in C#.NET...

Is there a performance hit for using the C++ API for CPLEX vs the C#.NET API? I'm wondering, because if possible I would like to avoid going across the managed/unmanaged boundary if possible, but I would like to be more informed about what is going on with the C#.NET API.

Does the C#.NET just implement the managed/unmanaged boundary to C++? Does it do it well?

Answer 1

The official ILOG documentation has this to say (emphasis mine):

Each call to a method of the API goes through a wrapping layer. This may result in a slight performance overhead while the model is created, compared to using the C++ API, depending on the number of API function calls. Since you call only few API functions to load and solve your model, the overhead is negligible in usual cases, but it may become important if you use the low-level Concert, CP Optimizer, or CPLEX® API for a complete model creation (for example, constructing a matrix line by line using INumExpr APIs or adding IConstraint objects one by one to an IModel using the API). It is therefore recommended to use the OPL language to model your problems whenever possible, and use only the low-level Concert APIs for the parts that need it (runtime additions, etc.).

Answer 2

We have been writing CPLEX + Concert code in C++ or C# and even some Java for about 18 years. These are many separate projects for many diverse customers. The preferred choice for many projects was originally C++ because we started before C# existed. Then C# was too new and wasn't widely trusted. But since about 2005, we have been preferring C# because it is easier to program in C#. We did one project in about 2004 where we had to write the CPLEX modelling code in C++, but had to use a C# database interface library (it was a complex project, and the reasoning behind those choices wouldn't make sense today). There was a definite overhead of using the interface between C# and C++; but those effects were much smaller than the differences between using different database access methods like ado.net vs odbc.

In practice, for most of our projects, the extra efficiency gains of using C++ are very small. Typically the created system will spend more than 90% (or 99%) of its time sat inside the CPLEX library calls. We have had many cases where the code in C++ or C# takes maybe 1 or 2 minutes to read and process the input data and create the CPLEX modelling variables and constraints, then 2-10 hours solving the problem inside cplex.solve(). So even if doing the processing outside of CPLEX in C++ instead of C# was ten times as fast, the overall timings aren't going to change a great deal.

If you are writing complex algorithmic stuff that builds and solves many (small) models (e.g. if you are doing column generation or LNS or similar) then there may be measurable benefits in C++, but probably not much in the wider picture.

I would go with whatever language you are more comfortable using. Getting the code right matters more than saving a few seconds.

Answer 3

Yes, C++ faster, check this:

• http://www.codeproject.com/Articles/253444/PInvoke-Performance
• http://msdn.microsoft.com/en-us/library/ky8kkddw.aspx

C# P/Invoke put some overhead for error handling, etc.