we could use our own bump allocator based on a known number of pages for the in-process stuff. this should improve performance and would prevent us from "tainting" the RSS as we could remove the known number of pages form the numbers we read in from /proc/self/statm.
this should work as most of our data structures, i.e. the IP<->index tree, as well as many hashes we use for interpretation later, only grow in size and can then be deleted in one step.