There seems to be a wrong apr pool used for processing certain MCMP commands. We should use short life span pools for immediate processing and server-lifetime pools only for really persistent configuration. In the current state, with 20+ tomcat workers (1 alias and 1 context each) and virtually no client requests, we could see slow, but steady growth of heap allocated memory.
TODO: Investigate the offending logic, make sure we ain't using long-lived pools for immediate processing.
Illustrative memory overview - with constant number of Tomcats: