These performance tests were run on both JIRA and Confluence clustering and required a large amount of hardware. We used the lowest end ancient 1 RU servers we could find to perform this proof of concept. You can see the performance increases for yourself in booth Confluence and JIRA when we went from one server to two.
Generally, we can tweak JIRA and Confluence single server instances to give excellent performance. For instance, we can run over 13 JIRA transactions per second with 50% read and 50% write. Combining that tweaking with the clustering would give even better performance increases yet.
There are many traditional ways to improve performance from HW to Networking to Apache which would all help of course. This document focused on Confluence itself.