Isn't that the point of your original question: you do not, in fact, know in advance the number of bytes you will need ?
I do know the full size, but I don't know the size of each vector. The data are organized as : <nbElements> elt1 elt2 ... <nbElements> elt1 elt2 elt3 ..., So, if I have 2 inner vectors, I'll have, ie : 5 1.2 3.3 2.5 4.6 0.2 2 0.5 5.6 nbElments are integers, but the other values have a known other type. Also, I know the number of "nbElements tags". So I know the number of inner vectors I'll need. I know also the number of values. But I don't know in advance how the values are spread in the vectors. I have to read the file.
If you have a way to move forward without getting stuck on this issue, finish your code, evaluate its performance and then -- if necessary run a profiler to figure out what actually needs improving.
Unfortunately, I don't have access to a profiler. But anyway, these classes seem to consume a lot of RAM compared to a solution based on simple pointers (I havn't compared to the offset_ptr<>, though...).