A memory allocator optimized for high concurrency and low memory consumption on modern multi-core and multi-processor systems. It includes special support for NUMA systems (allocating memory from the nearest NUMA node, for example).

It is about 60% faster than the allocator included with Microsoft Visual C++ and uses less than half the memory for small objects (< 256 bytes), leading to better cache utilization; this also makes it ideal for applications that allocate a huge number of small objects.
In many tests it is about 10% faster than the allocator included with Intel Thread Building Blocks.

It is implemented in C++ with a small amount of x86 Assembly and was recently converted to C++11.
Access to low-level synchronization and bit-manipulation instructions is done through compiler intrinsics.

The tests were done on a quad-core Intel Core i7 3770 CPU under Windows 8, 64 bit and Release mode. Because the CRT ultimately calls HeapAlloc, the version of Windows used is very important, the test application using the CRT allocator running much slower under Windows XP, which doesn't implement the Low Fragmentation Heap introduced with Windows 7.

The architecture and the implementation details are described in the following document:
Download architecture and implementation details (PDF)

The following documents present more implementation details:
Download internal data structures details (PDF)
Download bin location size details (PDF)