Erudil has an article in Linux Journal that mentions using NPH (non-parsed headers) to push 100's of megabytes of data to browsers over multi-week experiments. Look about 3/4ths of the way down the article.