MemoryManager causes a performance regression and should be replaced by a faster algorithm or, if we fail to find/implement one, just malloc and free, which still provides better performance than the current implementation.

Here are a couple of existing lock-free allocators I found:

• Streamflow (GitHub project). Edit: Implementation is POSIX-only.
  In case anyone's interested in the technical details, there's a paper describing the algorithm, including benchmarks: http://people.cs.vt.edu/~scschnei/papers/ismm06.pdf
• Lockless
• TCMalloc
• fc_malloc
• rpmalloc

Related issues: #2029, #3312, #3767