Kpm sometimes dies with a segmentation fault if more than one process is marked and the list is updated. It is unclear why this happens.
Kpm can consume a lot of CPU resources if the update period is too small. On the developer's system (AMD K6PR200) it uses approximatly 5 percent CPU time with an update period of one second. This depends heavily on the total number of processes (not just those displayed) and can be much higher than this under the right circumstances.
The %CPU number isn't accurate at very short update periods due to timer granularity.
The WCHAN field isn't correct if a process sleeps in a location outside those listed in System.map (for instance, in a kernel module). In these cases, the last entry in System.map will be shown (something like _end).