From my understanding of your problem: you have a index file that has millions of entries that dictate which file is to be read from, and the numbers somehow indicate the index into the file (quite possibly character positions or such.)
if that's the case, then i would first sort the index file itself by file number/name column, and then by the character/line indexing columns as required. then as long as the character/column indexes do not overlap within a single file, you should be able to process your such sorted index file and read the indicated files and ranges of whatever in sequential fashion.