TechPrufer
Get started
Meta
Meta logoMeta · Feature trail

Scalable Memory Allocation

3 posts · Jan 2011 – Mar 2026

What this trail covers

Meta's approach to memory allocation has evolved to address the challenges of high-throughput, low-fragmentation, and robust heap profiling in large-scale server applications. This includes the adoption and optimization of jemalloc, a memory allocator that provides fast allocation/deallocation, consistent RAM usage through low fragmentation, and critical heap profiling capabilities. Innovations have focused on improving thread caching, increasing mutex granularity, optimizing dirty page purging, and a renewed commitment to modernizing the codebase and evolving the allocator for the latest hardware and workloads, including improvements to the huge-page allocator, memory efficiency mechanisms, and AArch64 optimizations.

All Meta trails
  1. Investing in Infrastructure: Meta’s Renewed Commitment to jemallocThis post details Meta's renewed commitment to the jemalloc memory allocator. It outlines plans to reduce technical debt, modernize the codebase, and evolve jemalloc for new hardware and workloads. Specific areas of focus include improving the huge-page allocator (HPA) for better CPU efficiency with transparent hugepages (THP), enhancing memory efficiency through packing, caching, and purging mechanisms, and optimizing performance for the AArch64 (ARM64) platform. The post also emphasizes collaboration with the open-source community.
  2. Scaling memcache at FacebookThis post details Facebook's early adoption and scaling of memcached, starting in 2005. It highlights the critical role of memcache in enabling Facebook to scale to over a billion users by serving billions of requests per second. The post outlines specific infrastructure challenges faced due to the social graph's interconnectedness, large fan-out queries, and rapid product development. It also references a technical paper published in the 10th Usenix NSDI 2013 proceedings for deeper technical details.
  3. Scalable memory allocation using jemallocThis post details the core algorithms and data structures of jemalloc, including its segregation of small objects, careful size class selection, metadata overhead limits, minimization of active page set, and strategies to minimize lock contention through independent arenas and thread-specific caching. It also highlights Facebook-motivated innovations focused on improving speed through rewritten thread caching, increased mutex granularity, restructured synchronization logic to drop mutexes during system calls, rewritten dirty page purging, and a new red-black tree implementation. Introspection capabilities were enhanced with the `mallctl*()` API.