You know how the databases are clustered? Well, cluster 1 was running on older hardware with less memory. So over time it became the bottleneck. The web servers aren't partitioned by user, so they all felt cluster 1's bottleneck.
Added one new server to cluster 1, and things be flyin' now. And we still have 2 more servers to add.
Sorry about these last couple days of suck. I'll be better in the future about setting things up promptly.
Stats: 41.2 Mb/s; 563 connections/s