big transaction slows down over time - but disk seems almost unused

I've got a long-running, update-heavy transaction that increasingly slows
down the longer it runs. I would expect that behavior, if there was some
temp file creation going on. But monitoring vmstat over the life of the
transaction shows virtually zero disk activity. Instead, the system has
its CPU pegged the whole time.

So.... why the slowdown? Is it a MVCC thing? A side effect of calling
stored proceedures a couple hundred thousand times in a single
transaction? Or am I just doing something wrong?