mempool

A tcmalloc‑inspired memory pool for Linux with:

• 8 KB internal pages backed by mmap/munmap (alignment shim ensures 8 KB boundaries)
• Per-thread small-object caches to serve most allocations without locks
• Central caches per size class for batched recycle and load balancing
• 2-level Radix-tree page map for O(1) page→span lookups with lock-free read

---

Build

make

# defaults: 8 thr, 300k iters, sizes 8–256, keep 75%
make run-bench

# customize:
make run-bench NTHREADS=16 ITERS=400000 MIN=8 MAX=256 KEEP=75

Core sources :

• src/central_cache.cpp
• src/page_cache.cpp
• src/thread_cache.cpp

Tools

• tools/tb.cpp → build/bin/tb
• tools/allocator_tb.cpp → build/bin/allocator_tb

---

Benchmark: tb

Usage

./build/bin/tb <threads> <iters_per_thread> <min_size> <max_size> <keep_percent>

Examples

# Typical small‑object test (8–256 bytes)
./build/bin/tb 8 300000 8 256 75

# Include slightly larger small objects
./build/bin/tb 8 300000 8 512 75

# Scale threads
./build/bin/tb 16 400000 8 256 75

---

Features (inspired by tcmalloc)

• Three-layer design: ThreadCache (tc) → CentralCache (cc) → PageCache (pc), mirroring the simplified tcmalloc layout
• Per-thread caches (TLS) for small objects: tc handles requests up to ~256 KB before taking a slower path
• CentralCache uses hashed buckets with per-bucket locks, reducing global contention and enabling cross-thread rebalancing
• PageCache manages page-spans, performs split/merge (coalescing) of adjacent pages to reduce external fragmentation. (Pages default 8 KB here)
• Size classes & alignment: requests are rounded to the nearest class; batched refill/drain between tc↔cc amortizes locking
• Fast pointer→span lookup via a radix-tree page map (used to address MapObj→Span lookup bottlenecks)
• Large allocations (>~256 KB) bypass tc and take a separate path appropriate for big blocks