perf in the Linux kernel needs to process and count "callchains" (callstacks or call graphs) when reporting on profiles captured with perf record -g. Currently (as of August 2024) these are maintained in a kind of augmented red-black tree / radix tree with nodes that contain a singly-linked list of callstack entries (shared object pointer and instruction pointer pairs). Common prefixes in callstacks share nodes and split at the point where the callstacks differ.
When reading large perf.data files this design suffers from the usual linked list problem in that most of the execution time is eaten up by cache misses due to the random nature of following a list's ->next pointer. It's not unusual for 20-30% of startup time of perf report to be spent traversing these lists.