Multicore aware dump and restore would be a big win. I've hacked something up to do it in python queues and threads calling out to pg_dump. The drawback is that when I do it, I'm not getting a consistent, one transaction dump. But then usually, I'm shutting down the system to do an upgrade when I thwack it like this, so it's a reasonable tradeoff. With a fast set of drives, I can generally dump and reload with a concurrency of roughly O(#cores) in approximately the time it takes for the largest table, rather than the sum of the times for all the tables.

(there's a caveat there, and that's that my biggest tables tend to be toasted, so there's a decent processor load uncompressing the toasted bits. Also, Power law governs table size, and the biggest table is ~ 50+ % of the database, with ~ 100 tables. YMMV)

I wish there was a way to transparently parallelize operations over more than one core. I've got enough operations that are actually processor bound that I could benefit (some) from it. I could partition the database, but that's a management headache.