FinAlloc

custom memory allocator for quant infra

setup

1. clone
2. run makefile via make all

run

• run tests via make test
• executable via ./bin/finalloc
• benchmark via:

# per-thread pool, immediate alloc/free
./bin/allocBench --allocator=pool --threads=8 --iters=100000 --size=64

# lock-free pool, with a live set of 1024 (churn)
./bin/allocBench --allocator=lockfree --threads=8 --iters=200000 --size=64 --live=1024

# arena (per-thread). With --live>0 it does epoch resets every live/threads ops
./bin/allocBench --allocator=arena --threads=8 --iters=200000 --size=64

# baseline new/delete (immediate or churn with --live)
./bin/allocBench --allocator=new --threads=8 --iters=50000 --size=64

todos

1. arena allocator

• Multi-Chunk Growth Strategy [Dynamically allocates new chunks as the arena fills; allows for amortized linear growth with high locality]
• Custom construct(...) [API Supports object lifetime mgmt inside arena (T* ptr = arena.construct(...))]
• Memory Alignment + Padding Metadata [Tracks aligned offsets and optionally embeds headers to record allocation metadata]
• Guard Pages / Canary Support [For memory corruption detection (e.g., red zone under/overruns)]
• Thread-local Sub-Arenas [Arena-per-thread to avoid false sharing, tuned for CPU core affinity]
• Arena Group Manager [Shared chunk pool to reuse arenas between sessions (recycle slabs)]
• Journaling or Allocation Tracing Mode [For debugging perf regressions, e.g., log large allocations with source info]
• HugePage Support (Linux) [Backed by mmap with MAP_HUGETLB or madvise for better TLB performance]

2. pool allocator

• Lock-Free Free List [Use std::atomic<void*> with compare-and-swap to allow concurrent allocation/deallocation]
• Size-Class Bucketing [Support multiple object sizes by grouping into power-of-two buckets (like tcmalloc)]
• Thread-local Buffer Caches [Per-thread pools that reduce global contention and allocate in batches]
• Object Lifecycle Hooks [Optional callbacks for constructor/destructor on reuse, even for PODs]
• Zeroing or Poisoning Support [Debug mode wipes memory on alloc/dealloc to detect uninitialized accesses]
• Usage Metrics and Histograms [Track alloc counts, dealloc counts, high-water marks, fragmentation over time]
• Deferred Freeing / Quarantine [Introduce latency before freeing to detect use-after-free or memory races]

3. numa allocator

• Per-Node Slab Allocators [Maintain separate arena/pool per NUMA node (e.g., node 0 handles threads 0–15)]
• Thread Affinity Registry [Track each thread’s CPU/core → node mapping dynamically]
• Cross-Node Allocation Detection [Warn if a thread allocates from a remote NUMA node]
• Load-Balanced NUMA-Aware Arena Pools [Dynamically reallocate arenas across NUMA nodes to handle usage skew]
• Hardware Prefetch Hints [Use cache line prefetching (_mm_prefetch) to reduce stalls on frequent access patterns]
• NUMA-Integrated Memory Profiler [Visualize how much memory per NUMA node is in use, fragmented, idle]