I've been thinking about setting up a raw test run from EC2 hitting a couchdb instance several hundred times a second, just to see what happens. It won't be the best data for real-world performance, but I think it would still be valuable as a raw performance baseline.

Due to the nature of the CouchDB architecture, namely that views must be indexed, I would be more concerned with inserting data into CouchDB. CouchDB provides incremental replication, which means only documents that have changed need to be re-indexed. However, if you have a lot of initial data that needs to be imported, you could see a lot of latency there.

I assume you mean EC2?

I do now =)