Hacker News Reader

Comment by cperciva | Hacker News Reader

> Memory management is important even for short-lived programs. It would bring burden to the OS if you invoke this kind of "short-lived, memory-management-free" programs multiple times.

All the memory is freed when the process exits. Why does it matter if you run the program multiple times?

View on HN

It will put more pressure on the allocator when running, doing that a lot will likely have some kind of cumulative consequence down the line. I know from experience [0] that even reusing allocated memory rather than bouncing it back to malloc can have dramatic effects.

[0] https://gitlab.com/sifoo/snigl/blob/master/src/snigl/pool.h

Sure, but if you're not freeing memory you don't need a real malloc. You allocation function can be just "top += allocsize".

Yes, that's commonly known as "arena allocation" and has been used for decades in programs that have well-defined phases that create lots of connected objects with the same lifetime, like compiler passes, RPC serialization layers, or query planners.

Wait, doesn't it do the opposite? The big thing that "pressures" an allocator is fragmentation, which you don't get at all if you never free.

With no freeing, memory used is the sum of all allocations. With freeing, it is max of the allocation for active objects at any given time. The former can ask for much more memory from the OS for some programs that generate a lot of temp garbage. Though this probably doesn’t matter for a C compiler - unless it is required to compile megabytes of generated code!

An allocation you don't have to free is, at the margin, a simple pointer increment. It's the need to free memory that makes allocators expensive, not the amount of memory they allocate. This is why performance-sensitive programs use pools, arenas, and other preallocation strategies.

What I was getting at is that non-freeing programs would use much more memory. This doesn't increase the "pressure" on the allocator for such a program but does increase the pressure on the OS (that has to handle memory requests for many concurrently running programs).

Malloc doesn't interact with the kernel at all. The kernel sees pages, not the data structures that malloc manages. The kernel doesn't even know whether you free()'d the memory by the time the process exits.

There is exactly zero difference from the operating system's perspective between freeing and not freeing the memory before program termination (except that one might have a higher peak memory usage).

The classic implementations of several standard Unix utilities deliberately never deallocated memory for performance reasons. If there was a problem with this practice, I would expect Bell Labs--of all places--to know not to do it.

>There is exactly zero difference from the operating system's perspective (...)

Close, but not quite: in case of larger allocations, malloc() tends to use mmap( , , , | MAP_ANON) rather than brk() to request memory from the OS. For example, the glibc's malloc() uses mmap() when requested size exceeds MMAP_THRESHOLD, which is 128kB by default.

The mmap() approach gives large, continuous memory blocks that can also be easily free()'d via munmap() with little to no bookkeeping needed[1] - not being subject to the same fragmentation woes as memory allocated via brk() - as long as your address space is significantly larger than allocated memory.

That aside, I fully agree with the author of 9cc.

[1] in fact a simplistic libc memory allocator could directly wrap malloc() around mmap(), free() around munmap() and realloc() around mrealloc(), leveraging the in-kernel allocator at the cost of one syscall with at least two context switches at each call - i.e., slowww

Using mmap/munmap isn't quite that trivial: You need to remember how large each allocation is in order to know how many pages to unmap.

An implementation certainly could return memory to the OS on calls to free(), but to my knowledge none of the widely used implementations do so. (I would be interested to learn of counterexamples!)

jemalloc can use madvise(MADV_FREE) to return pages to the OS.

Did I mention kernels anywhere? I simply shared my experience.

If that doesn't fit neatly into your delusions it's really not my problem.

Reply on news.ycombinator.com