I think this is an extremely difficult question. Urgency can be measured in months.
I work in a large company that uses std::string as its string choice. The application we code is a database application, and it often does hundreds of thousands of appends to a single string for a single report (The format we output is XML). The data in a report can be anywhere from 1k, to about 1gig. My group has developed a wrapper around this template class, but when we pass a std::string to library, we can not pass our own class, so this problem needs to somehow be handled in the std::string class itself. Maybe via the allocator object.
The problem is, these appends slow us down because of all the allocation and string copying that must be done each append. If we could somehow have std::string store all the appends in a linked-list, and then combine it once, when we need it, this would be an ideal solution. Any speed-ups that can be provided could be helpful though.
We need this in the std::string itself because this is the only object we can pass to other libraries that need to read or write to the string. While we could make sure the string was concatenated before passing it outside of our code, this is very dangerous and tedious since it's easy to forget.
Allocating more space than needed would help, but wouldn't be an ideal solution since it would hamper performance. The large range in amount of data that can be returned makes it difficult to guess how much would be good since a single append is generally anywhere from 2 bytes to 400 bytes.
I wasn't able to understand how the allocator class could be overridden since we need to combine all the data when it is needed.