Is accessing data in the heap faster than from the stack?

Is accessing data in the heap faster than from the stack?

Not inherently... on every architecture I've ever worked on, all the process "memory" can be expected to operate at the same set of speeds, based on which level of CPU cache / RAM / swap file is holding the current data, and any hardware-level synchronisation delays that operations on that memory may trigger to make it visible to other processes, incorporate other processes'/CPU (core)'s changes etc..

The OS (which is responsible for page faulting / swapping), and the hardware (CPU) trapping on accesses to not-yet-accessed or swapped-out pages, would not even be tracking which pages are "global" vs "stack" vs "heap"... a memory page is a memory page.

While the global vs stack vs heap usage to which memory is put is unknown to the OS and hardware, and all are backed by the same type of memory with the same performance characteristics, there are other subtle considerations (described in detail after this list):

• allocation - time the program spends "allocating" and "deallocating" memory, including occasional sbrk (or similar) virtual address allocation as the heap usage grows
• access - differences in the CPU instructions used by the program to access globals vs stack vs heap, and extra indirection via a runtime pointer when using heap-based data,
• layout - certain data structures ("containers" / "collections") are more cache-friendly (hence faster), while general purpose implementations of some require heap allocations and may be less cache friendly.

Allocation and deallocation

For global data (including C++ namespace data members), the virtual address will typically be calculated and hardcoded at compile time (possibly in absolute terms, or as an offset from a segment register; occasionally it may need tweaking as the process is loaded by the OS).

For stack-based data, the stack-pointer-register-relative address can also be calculated and hardcoded at compile time. Then the stack-pointer-register may be adjusted by the total size of function arguments, local variables, return addresses and saved CPU registers as the function is entered and returns (i.e. at runtime). Adding more stack-based variables will just change the total size used to adjust the stack-pointer-register, rather than having an increasingly detrimental effect.

Both of the above are effectively free of runtime allocation/deallocation overhead, while heap based overheads are very real and may be significant for some applications...

For heap-based data, a runtime heap allocation library must consult and update its internal data structures to track which parts of the block(s) aka pool(s) of heap memory it manages are associated with specific pointers the library has provided to the application, until the application frees or deletes the memory. If there is insufficient virtual address space for heap memory, it may need to call an OS function like sbrk to request more memory (Linux may also call mmap to create backing memory for large memory requests, then unmap that memory on free/delete).

Access

Because the absolute virtual address, or a segment- or stack-pointer-register-relative address can be calculated at compile time for global and stack based data, runtime access is very fast.

With heap hosted data, the program has to access the data via a runtime-determined pointer holding the virtual memory address on the heap, sometimes with an offset from the pointer to a specific data member applied at runtime. That may take a little longer on some architectures.

For the heap access, both the pointer and the heap memory must be in registers for the data to be accessible (so there's more demand on CPU caches, and at scale - more cache misses/faulting overheads).

Note: these costs are often insignificant - not even worth a look or second thought unless you're writing something where latency or throughput are enormously important.

Layout

If successive lines of your source code list global variables, they'll be arranged in adjacent memory locations (albeit with possible padding for alignment purposes). The same is true for stack-based variables listed in the same function. This is great: if you have X bytes of data, you might well find that - for N-byte cache lines - they're packed nicely into memory that can be accessed using X/N or X/N + 1 cache lines. It's quite likely that the other nearby stack content - function arguments, return addresses etc. will be needed by your program around the same time, so the caching is very efficient.

When you use heap based memory, successive calls to the heap allocation library can easily return pointers to memory in different cache lines, especially if the allocation size differs a fair bit (e.g. a three byte allocation followed by a 13 byte allocation) or if there's already been a lot of allocation and deallocation (causing "fragmentation"). This means when you go to access a bunch of small heap-allocated memory, at worst you may need to fault in as many cache lines (in addition to needing to load the memory containing your pointers to the heap). The heap-allocated memory won't share cache lines with your stack-allocated data - no synergies there.

Additionally, the C++ Standard Library doesn't provide more complex data structures - like linked lists, balanced binary trees or hash tables - designed for use in stack-based memory. So, when using the stack programmers tend to do what they can with arrays, which are contiguous in memory, even if it means a little brute-force searching. The cache-efficiency may well make this better overall than heap based data containers where the elements are spread across more cache lines. Of course, stack usage doesn't scale to large numbers of elements, and - without at least a backup option of using heap - creates programs that stop working if given more data to process than expected.

Discussion of your example program

In your example you're contrasting a global variable with a function-local (stack/automatic) variable... there's no heap involved. Heap memory comes from new or malloc/realloc. For heap memory, the performance issue worth noting is that the application itself is keeping track of how much memory is in use at which addresses - the records of all that take some time to update as pointers to memory are handed out by new/malloc/realloc, and some more time to update as the pointers are deleted or freed.

For global variables, the allocation of memory may effectively be done at compile time, while for stack based variables there's normally a stack pointer that's incremented by the compile-time-calculated sum of the sizes of local variables (and some housekeeping data) each time a function is called. So, when main() is called there may be some time to modify the stack pointer, but it's probably just being modified by a different amount rather than not modified if there's no buffer and modified if there is, so there's no difference in runtime performance at all.

Note

I omit some boring and largely irrelevant details above. For example, some CPUs use "windows" of registers to save the state of one function as they enter a call to another function; some function state will be saved in registers rather than on the stack; some function arguments will be passed in registers rather than on the stack; not all Operating Systems use virtual addressing; some non-PC-grade hardware may have more complex memory architecture with different implications....

Quoting from Jeff Hill's answer:

The stack is faster because the access pattern makes it trivial to allocate and deallocate memory from it (a pointer/integer is simply incremented or decremented), while the heap has much more complex bookkeeping involved in an allocation or free. Also, each byte in the stack tends to be reused very frequently which means it tends to be mapped to the processor's cache, making it very fast. Another performance hit for the heap is that the heap, being mostly a global resource, typically has to be multi-threading safe, i.e. each allocation and deallocation needs to be - typically - synchronized with "all" other heap accesses in the program.

There is blog post available on this topic stack-allocation-vs-heap-allocation-performance-benchmark Which shows the allocation strategies benchmark. Test is written in C and performs compare between pure allocation attempts, and allocation with memory init. At different total data sizes, number of loops are performed and time is measured. Each allocation consists of 10 different alloc/init/free blocks with different sizes (total size shown in charts).

Test are run on Intel(R) Core(TM) i7-6600U CPU, Linux 64 bit, 4.15.0-50-generic, Spectre and Meltdown patches disabled.

With out init:

With init:

In the result we see that there is significant difference in pure allocations with out data init. The stack is faster than heap, but take a note that loop count is ultra high.

When allocated data is being processed, the gap between stack & heap performance seems to reduce. At 1M malloc/init/free (or stack alloc) loops with 10 allocation attempts at each loop, stack is only 8% ahead of heap in terms of total time.