Cost of static memory allocation vs dynamic memory allocation in C

Question

I am very interested to know what is the preferred method of memory allocation static vs dynamic is good for performance (e.g., running time) when you know the

Answer 1

True static allocations (globals, and local variables marked static) are glued into sections and loaded along with the rest of the section at load time (execve) using mmap. They're essentially free, already covered by the cost of loading the program.

Automatic arrays with statically known sizes can be glued together with other local variables and allocated by adjusting the stack pointer. That's essentially one integer subtraction (the stack grows downwards) or something close to 1 ns on modern processors.

Variable length arrays can't be glued to other variables so they should cost about 1 ns each.

I tried measuring mallocs with different sizes in a single threaded process and I got this, which would imply that malloc+free pairs cost about 50-100 times as much as stack variables for allocations <16MiB. After that, it spikes upwards (32MiB and 64MiB both cost about a hundred times as much as allocations <=16MiB do):

These are just the costs of obtaining the pointer though. Actual access may cause page faults and thereby further increase the cost of the memory block. The page faults should be extremely rare with stack allocations, but likely with first-time accesses to dynamically allocated memory.

Additionally, many small dynamic allocations will put quite a strain on caches as your memory will be fragmented. (You can reduce this fragmentation by using memory pools, either your own or the obstack functionality available as part of glibc.)