Memory Allocators: malloc vs. tcmalloc vs. jemalloc — What Really Happens

Systems Programming Deep Dive — Issue 1

Most services run on the allocator that shipped with glibc and never think about it. That’s fine until your allocation-heavy microservice starts spending 40% of wall time in _int_malloc and your on-call rotation gets interesting.

---

The Problem Nobody Tells You About

glibc’s allocator — ptmalloc2, unchanged since 2006 — was designed for single-threaded workloads and patched for threading later. The seams show. Each arena is a coarse-grained lock over a linked list of chunks. Under moderate thread contention, perf top starts looking like this:

  37.12%  libc.so  [.] _int_malloc
  14.88%  libc.so  [.] _int_free
   8.43%  kernel   [k] futex_wait_queue_me

That futex_wait line is threads sleeping on arena locks. You’re paying context switch cost on every allocation path, and the faster your threads allocate, the worse it gets.

---