Hello, I have been developing a piece of code using various boost libraries, but I have to pause because the compilation has become unwieldy. It's a 50kloc prgm, which takes about 30min to compile on a pentium 3.5GHz HT and 1GB of memory (gcc 3.4 on linux). The binary, using the static linking with 2 boost libraries (program-options, filesystem) yield a 380M binary. Furthermore, it doesn't compile on msvc 7.1 (c1063, compiler stack overflow), and I had to stop mingw after one hour.... The prgm handles the parsing, storage, and serialization of a big raw string. This string is organized into fragments, about 30 each fragment could be occuring 0 or more times. each fragment contain on average 30 fields. each fragment has little in common with other fragments. my solution was to store the string as a vector of a type Fragment described by typedef variant<fragment1, fragmen2, ..., fragment30> Fragment. typedef vector<Fragment> string_storage. the parsing is handled using a grammar for each fragment, and the grammars are split between the definition and instantiation (manual explicit), as suggested in the FAQ. Most fields have a unique functor which is used as a semantic action to store the data found in the fragment. So this all scheme looked darn good when I had only a couple fragments, and it was working very well too! But now, my computer threatens to go strike whenever I type bjam on the command line. I don't have a clear question, besides the simple 'how well does variant scale'?, or 'does that seem like a not so clever design?'. The variant page mentions an optimization opportunity with 'has_nothrow_copy', but I can't find a good example. Could that help? Thanks, Matt