My Problem

Process memory usage increases over weeks
Receives data from multiple sockets, processes, exports over sockets
~40 Threads active
~4-5 GiB of data in process concurrently
Very many allocations/deallocations, containers and small objects

Idea 1: profile

Try valgrind/massif
Try gperftools/tcmalloc heapprofile
to slow to see any results
>12h for some minutes of playback data

:-(

Idea2 2: counting_memory_resouce

count memory allocations, deallocations, usage
switch to pmr containers and allocate_shared
search for possible culprits

Demo Time (2)

demo2.cpp

Notes

Simple to add statistics
Memory Pools help
- avoid new/delete calls
- increase data locality
- but do not release memory until the pool is destroyed

Results for my Problem

One counting_memory_resouce per software Component
- see which component are most active
And one per major data type
- identify which data uses most memory
found several 100MiB unused data structures
memory pools seem to reduce memory fragmentation