There are no documented architecture about scalable usergrid deployment. You need to configure your own deployment based on your requirements. Some samples can be found on the internet, this presentation helped me to configure our usergrid installation: http://events.linuxfoundation.org/sites/events/files/slides/Intro-To-Usergrid%20-%20ApacheCon%20EU%202014.pdf (pages 47-48).
And here is my deployment strategy: All the components (tomcat, C*, es) are java applications, so putting them on to the same machine will be expensive on RAM. So, separate the layers, and scale them independently. For example, if your application chokes on incoming user connections, just scale up tomcat cluster (behind a LB probably). Spend time on configuring Cassandra, and don't stick to the default values - your data will be there and you don't want to lose it.