TCmalloc is used in MacrayKV as its underlying memory allocator. Here is some notes on TCmalloc.
tcmalloc motivation:
1. 50 nanoseconds of TCMalloc v.s 300 nanoseconds of glibc 2.3 malloc( which is called ptmalloc2 )
2. reduce lock contention for multi-threaded program.
3. space-efficient representation of small objects (<= 32K)
usage
link with -ltcmalloc
overview
1. small allocations are satisfied from thread-local cache.
2. objects are moved from a central object pool to thread-local cache as needed
3. periodic garbage collections are run to migrate object from thread-local cache back to central pool
4. large objects (>32K) are directly allocated from the central heap using a page-level(4K aligned) allocator.
small object allocation
1. the 32K space is divided into about 170 size-class. gaps between classes are vary from 8 to ~2K
2. direct allocate from local > direct allocate from central > allocate from sys and then allocate from central
large object allocation
1. a large object size (> 32K) is rounded up to a page size
http://goog-perftools.sourceforge.net/doc/tcmalloc.html