Matching addresses is a time consuming process. Everything is matched with everything, and that makes for a lot of work. It might help to filter the items to match by first splitting the address into seperate words, and then calculating a Jaccard index for each address combination. Only those that have a sufficiently high Jaccard score would need to be processed further. Perhaps a speed increase would be possible this way. Currently, processing 90k addressing takes about 7 hours on a quad core machine.