[WORK-IN-PROGRESS] np_alloc

• yet-another, per-thread memory alloc/pool for c++
• code from scratch, some idea from written commerical code by myself in 2009.

toolset

• c++14/gsl/x64
• consideration for linux, not yet buildable

feature

• lock-free, per-size, per-thread memory pool
• wait-free most of times except fetching from global pool

api

void*    np_alloc(size_t bytes);
void*    np_alloc(size_t bytes, const char file[], int line);
void     np_free(void * ptr);

usage & test code

• see unittest.cpp
• see core source

pros / cons

• (+) no lock to get new memory so fast
• (+) improved cache locality due to pre-allocated by memory page for local thread
• (-) using lots of memory but can be optimizing per project

todo

garbage collecting process from local thread pool to global pool

• it will saves memory but worse cache localtity due to spreading out memory fragments across threads

benchmark

• just for initial lame benchmarks. code are here
• methods: allocate up to max allocate count and deallocate until all freed and repeat this

thread [28b0]: global pool created. pool=[2042de1a4b0]
malloc {
         random_multithread (500000) - 5.11385s
         random_singlethread (5000000) - 1.43155s
         small_multithread (500000) - 0.733916s
         small_singlethread (5000000) - 0.281047s
         big_multithread (500000) - 3.08501s
         big_singlethread (5000000) - 1.76247s
}
np_alloc {
         random_multithread (500000) - 1.2951s
         random_singlethread (5000000) - 0.229591s
         small_multithread (500000) - 0.681932s
         small_singlethread (5000000) - 0.0977364s
         big_multithread (500000) - 0.826968s
         big_singlethread (5000000) - 0.232066s
}
thread [3a84]: global pool destroying. pool=[2042de1a4b0]
[3a84] ~global_pool() this=[2042de1a4b0]
Press any key to continue . . .

• windows 10 x64 / xeon x5570 2 processor / 8 core 16 thread
• env|50 threads|500k iteration per thread|10000 max alloction per thread

• env|1 thread|5000k iteration per thread|10000 max alloction per thread