Matthias Troyer wrote:

After a few tweaks - which I will report later - I got serialization to compile on my installation of MacOS X.

That is very good news - it's a platform I never tried. Please try and run the bjam test suite as well.

I have a question that does not seem to be addressed in the documentation: What is the difference betweem save and save_override or > load and load_override?

demo_fast_binary_archive.cpp uses save_override and load_override but I could not find it documented?

Making a new archive class is described in the section "New Archives - Implementation". The model for an archive class is: ///////////////////////////////////////////////////////////////////////// // class trivial_iarchive - read serialized objects from a input text stream class trivial_iarchive : public boost::archive::common_iarchive<trivial_iarchive> { // permit serialization system priledged access to permit // implementation of inline templates for maximum speed. friend class boost::serialization::load_access; // member template for loading primitive types. // Override for any types/templates that special treatment template<class T> void load(T & t); public: ////////////////////////////////////////////////////////// // public interface used by programs that use the // serialization library // the >> operator template<class T> trivial_iarchive & operator>>(T & t){ boost::serialization::load(* This(), t); return * This(); } // archives are expected to support this function void load_binary(void *address, size_t count); }; So in theory one has to implement the two undefined functions. For each primitive data type, int, float, etc. or any type designated as primitive define a "load" template for function to read them from the file. For each data type that you want to treat specially - rather than allowing the serialization library to do the job - define an overload for the >> operator. That should be all there is too it. The archive classes included in the library are much more elaborate for a number of reasons. Among others these are: a) portability to platforms which fail to support Partial Template Function Ordering. This is what make the load_override functions necessary so that we can apply the "extra argument" work around to simulate PTFO for these platforms. They are not part of an public interface but rather internal artifacts of the implementation of these archives. b) the desire that the archives be templates so that they could work with both char and wchar_t i/o c) the desire to factor out common code for text primitives so that it wouldn't have to be duplicated for xml and text archives.

I can implement an optimized archive along the lines of demo_fast_binary_archive.cpp, but extended to also support std::vector, std::valarray, ...

Before doing that, I would suggest making a test application with the standard binary archive and try to get some baseline timings. I would be curious to know the results of this. Remember that the point of demo_fast_binary_archive.cpp isn't so much to propose a faster archive ( though it may be ) but really to illustrate how one might derived from one of existing archives to alter its behavior. Robert Ramey