[0] https://doc.rust-lang.org/beta/rustc/target-tier-policy.html [1] https://doc.rust-lang.org/rustc/targets/custom.html

After some pondering, and reading the rust documentation, I came to the conclusion that te difference is this: reserve() will grow the underlaying memory area to the next increment, or more than one increment, while reserve_exact() will only grow the underlaying memory area to the next increment, but no more than that.

Eg, if grow strategy is powers of two, and we are at pow(2), then reserve() may skip from pow(2) to pow(4), but reserve_exact() would be constrained to pow(3) as the next increment.

Or so i read the documentation. Hopefully someone can confirm?

The C++ one, however, will not reserve more than you ask for (in the case that you reserve greater than the current capacity). It's an exact reservation in the rust sense.

> reserve() will grow the underlaying memory area to the next increment, or more than one increment, while reserve_exact() will only grow the underlaying memory area to the next increment, but no more than that

No, not quite. Reserve will request as many increments as it needs, and reserve_exact will request the exact total capacity it needs.

Where the docs get confusing, is that the allocator also has a say here. In either case, if you ask for 21 items, and the allocator decides it prefers to give you a full page of memory that can contain, say, 32 items... then the Vec will use all the capacity returned by the allocator.

Of course, this can change in the future; in particular, the entire allocator API is still unstable and likely won't stabilize any time soon.

We could, as that comment suggests, check the slice and see if there's enough room for more than our chosen capacity. We could also debug_assert that it's not less room, 'cos the Allocator promised it would be big enough. I dunno if that's worthwhile.

says

> Increase the capacity of the vector (the total number of elements that the vector can hold without requiring reallocation) to a value that's greater or equal to new_cap.

I belive that the behaviour of reserve() is implementation defined.

If you don't offer Vec::reserve_exact then people who needed that run out of RAM and will dub your stdlib garbage. If you don't offer Vec::reserve as we've seen C++ programmers will say "Skill issue" whenever a noob gets awful performance as a result. So, it's an easy choice.

It would be nice if this were true but AFAIK the memory allocator interface is busted - Rust inherits the malloc-style from C/C++ which doesn’t permit the allocator to tell the application “you asked for 128 bytes but I gave you an allocation for 256”. The alloc method just returns a naked u8 pointer.

But the (not yet stable) Allocator::allocate returns Result<NonNull<[u8]>, AllocError> --- that is, either a slice of bytes OR a failure.

Vec actually relies on Allocator not GlobalAlloc (it's part of the standard library so it's allowed to use unstable features)

So that interface is allowed to say you asked for 128 bytes but here's 256. Or, more likely, you asked for 940 bytes, but here's 1024. So if you were trying to make a Vec<TwentyByteThing> and Vec::with_capacity(47) it would be practical to adjust this so that when the allocator has 1024 bytes available but not 940 we get back a Vec with capacity 51 not 47.

By contrast reserve will allocate space for the extra elements following the growth strategy. If you reserve(100) on an empty Vec the allocation will be able to actually insert 128 (assuming the growth strategy is pow(n))

Vec::reserve(100) on an empty Vec will give you capacity 100, not 128 even though our amortization is indeed doubling.

The rules go roughly like this, suppose length is L, present capacity is C, reserve(N):

1. L + N < C ? Enough capacity already, we're done, return

2. L + N <= C * 2 ? Ordinary doubling, grow to capacity C * 2

3. Otherwise, try to grow to L + N

This means we can grow any amount more quickly than the amortized growth strategy or at the same speed - but never less quickly. We can go 100, 250, 600, 1300 and we can go 100, 200, 400, 800, 1600 - but we can''t do 100, 150, 200, 250, 300, 350, 400, 450, 500...

The higher level the language, the less interest there is to manually manage memory. It is just something to offload to the gc/runtime/etc.

So, i think this is a no-go. The market wont accept it.

If you really want to dumb down what I’m suggesting, it’s is tantamount to blade servers with a better backplane, treating the box as a single machine instead of a cluster. If IPC replaces a lot of the RPC, you kick the Amdahl’s Law can down the road at least a couple of process cycles before we have to think of more clever things to do.

We didn’t have any of the right tooling in place fifteen years ago when this problem first started to be unavoidable, but is now within reach, if not in fact extant.

First, amdahl's law just says that the non parallel parts of a program become more of a bottleneck as the parallel parts are split up more. It's trivial and obvious, it has nothing to do with being able to scale to more cores because it has nothing to do with how much can be parallelized.

Second in your other comment, there is nothing special about "rust having the semantics" for NUMA. People have been programming NUMA machines since they existed (obviously). NUMA just means that some memory addresses are local and some are not local. If you want things to be fast you need to use the addresses that are local as much as possible.

Don't assume anything in shells except POSIX sh.

Unfortunately there are people who think Linux is the only way.

If the underlying resource can not behave like a class, it would be better to expose a free function style api, eg:

    Handle h = ResourceGet();
    ResourceDoSomething(h);
    h.release();

Should it not be std::array<>?

sp::primitive_array holds memory following the Arrow specification, which can be operated upon in place from e.g. ArrowCpp, PyArrow, etc.