list.pop(0) vs. collections.dequeue
showell30 at yahoo.com
Sat Jan 23 00:17:21 CET 2010
On Jan 22, 2:54 pm, Dave Angel <da... at ieee.org> wrote:
> Steve Howell wrote:
> > On Jan 22, 12:40 pm, Christian Heimes <li... at cheimes.de> wrote:
> >> Steve Howell wrote:
> >>> Is that really true in CPython? It seems like you could advance the
> >>> pointer instead of shifting all the elements. It would create some
> >>> nuances with respect to reclaiming the memory, but it seems like an
> >>> easy way to make lists perform better under a pretty reasonable use
> >>> case.
> >>> Does anybody know off the top of their head if the "have-to-be-shifted-
> >>> by-one" warning is actually valid?
> >> Why do you think the documentation has such obvious errors?
> > I wasn't making any assumptions, hence the question mark. The Python
> > docs are very good, but even the best of projects make advances that
> > aren't reflected in the docs.
> >> CPython's list type uses an array of pointers to store its members. The
> >> type is optimized for the most common list operations in Python:
> >> iteration and appending. Python code rarely changes the head or middle
> >> of a list. The dequeue type is an optimized data structure for popping
> >> and inserting data at both ends.
> > I disagree that Python code rarely pops elements off the top of a
> > list. There are perfectly valid use cases for wanting a list over a
> > dequeue without having to pay O(N) for pop(0). Maybe we are just
> > quibbling over the meaning of "rarely."
> >> When you list.pop() or list.insert() the pointers in the internal array
> >> must be shifted. appending is much faster because the internal array is
> >> overallocated meaning it contains free slots at the tail of the data
> >> structure. A linked list of pointers requires more memory and iteration
> >> is slower since since it can't utilize the CPU's cache as good as an array.
> > I am not proposing a linked list of pointers. I am wondering about
> > something like this:
> > p =p;
> > (and then reclaim p as free memory, I already said I understood
> > that was the tricky bit)
> > The pointer arithmetic for accessing each element would still work in O
> > (1), right?
> I think it was Dijkstr (sp?) who said you can accomplish anything with
> just one more level of indirection. Clearly each attribute (variable)
> that has a binding to a given list must point to a fixed piece of
> memory, that cannot safely be moved, because there's no efficient way to
> find all the attributes. That fixed piece is the list object, and I
> expect it's 16 or 20 bytes, on a 32bit implementation. It must in turn
> point to the actual malloc'ed block that contains pointers to all the
> elements of the list. So that block will be 4*n where n is the number
> of reserved cells, at least as large as len(). This is the block where
> copying takes place when you insert or delete from the beginning.
The indirection is already in Python, as it (at least appears to me)
that everything is deferenced off of an ob_item pointer:
> The list object must contain a pointer to the beginning of this block,
> or it wouldn't be able to free() it later. So you'd be suggesting a
> second pointer that actually points to the current 0th pointer. And a
> pop would simply increment this second pointer.
Yes, ob_item would point to the 0th element pointer and pop would
simply increment it.
The additional bookkeeping would be the original pointer.
> Such an approach might be worthwhile if you expect lots of pops and
> pushes, with a minimal net change. But of course each time you did a
> pop, you'd have to decide whether it was time to normalize/compact the
Yes, and that of course is the tricky bit.
> As you say, reclaiming the 0th element of this block to the memory pool
> would be tricky.
I should clarify that a bit. Reclaiming the 0th element *cheaply* is
tricky, unless you want to rewrite the memory manager. But I also
think you can, of course, defer reclaiming the element.
> Doubly so, because 1) the C memory allocator has no
> such notion as resizing the beginning of a block. and 2) it would have
> nothing useful to do with the 4 bytes freed. The minimum allocated
> block in Python is probably 16 bytes of actual address space. I'd guess
> that's 4 bytes for malloc's overhead, and 8 bytes for the minimum object
> header, and 4 bytes for data. To check me, try:
> >>> a = 5.3
> >>> b = 49.6
> >>> id(a), id(b)
> (11074136, 11074152)
> Anyway, this could be done, where once the two pointers get some
> distance apart, you do a realloc, and copy everything. But of course
> you'd want to build some hysteresis into it, to avoid thrashing.
> , but
There wouldn't be any additional thrashing beyond what happens now.
You'd simply avoid the first N-1 memmoves of up to kN bytes at the
cost of not reclaiming those k(N-1) bytes right away. I suppose it's
a more "bursty" penalty you'd be paying, but the peak of the "bursty"
curve is no wider than the "constant" curve, it's just N times
> There wouldn't be much of a performance hit, but it would increase every
> list size by 4 bytes minimum. So I doubt it would be a list replacement.
I don't think that's the reason people would oppose this, although you
are true about the penalty. Memory's cheap, you'd need about a
quarter million lists just to fill up a meg.
CPU cycles, on the other hand, are expensive, as users' demand for
responsive programs seems to do a better job of keeping up with
I'd also argue that the memory you keep around to avoid unnecessary
memmoves() will almost always be dwarfed by the memory used by the
list elements themselves.
More information about the Python-list