As for distributed computing, there is a distributed computing climatology study under way: http://climateprediction.net/ . It runs on both the BOINC system (same as Seti@Home uses) or under it's own system. It takes about 600 hours/model run on an Athlon XP 3000 or Pentium 4 3.0GHz.

As for model runs such as the GFS, the problem is timeliness. You can send out a packet of data for analysis, but can you get it back in a timely manner? It might be a couple days before you get that little piece back. So, distributed computing is fine for long term, distinct-solution systems, but it doesn't really work for something like a 6-hourly model that every layer of data builds on the previous one.

--RC