I am attempting a project that uses DLLs as a means of having a dynamic object factory. In other words, I have a bunch of concrete classes derived from a single abstract class, but I do not know what actual class I am using until run-time. Instead of creating a large switch statement that has to be modified each time I make a new derived class, I create a new DLL for each derived class. Each DLL provides an interface function that when invoked returns an instance of the derived object. So, instead of a large switch statement, I just load the DLL based on a file name and get the interface function every time.

The heart of the problem is that I wanted to test the overhead incurred when using the DLL objects instead of the hardcoded ones. To do so, I threw in a large for loop into one of the object functions that called pow(rand(), 3) 50000 times. I then profiled the two versions. To my surprise, using the AQ time profiler, after 500 (or 500*50000 = 2.5M rand calls) tests, the total time that the program spent in the rand() call when using DLLs was more than twice the time spent in rand() when not using DLLs.

Is there any easy explanation to this? I am very new to DLL usage, but I was not aware of any reason why calling rand from with a DLL created object would incur so much more overhead than a call from a non-DLL object. Perhaps my testing methodology is wrong? I am currently using VC++ 6.0 with the STLPort 4.6.2 libraries (since the MS STL does not support passing STL objects from a program to a DLL function, which is a requrement for some of the functoins in my objects).

Any help would be greatly appreciated!