Hacker News Reader

Comment by herf | Hacker News Reader

In some cases you can benefit by "struct of arrays" instead of "array of structs". This benefits packing, enables SIMD, and usually improves persistence speed too. Database people call this a "column store". So if you abstract it a little you can do per-column compression that is really hard to do with vanilla structs.

One thing I learned with Picasa (which is basically an in-memory database): if you're storing a lot of data in RAM, sometimes you should think of the data structures more as a database (with copy in/out, rather than reference). In multithreaded cases, this kind of approach unexpectedly gets you more parallelism and speed.

View on HN

Column store is really just making sure values for a column are continuous in memory, instead of rows as it is done usually. Basic premise is that modern CPU's can chew through continuous memory locations at light speed, with single core being capable of scanning a gigabyte of RAM in few milliseconds[0].

This is extremely beneficial when your tables have large number of columns, but you only want to do computations on a few of them. Columnar stores usually implemented using dictionaries which allow for efficient duplicate handling and blazing fast lookups. Often those dictionaries also contain run length optimizations, so values 1 through 100 are only stored using two values start-range 1 and end-range 100.

All in all, column stores are efficient at analytic workload, but struggle when a lot of inserts and updates need to happen.

[0] Using prefetch, properly aligned data, and likely other optimizations I am not thinking of.

but struggle when a lot of inserts and updates need to happen.

I ran across this with Redis recently. Redis has an internal data structure called ziplist that's a doubly linked list with no pointers. It just uses byte offsets to the next element, so all elements are in one contiguous block of memory. It's great because you can store a lot of values compactly without needing 1 to 5 pointers (8 to 40 bytes) overhead per element in your data structure.

But, a problem shows up when you want to insert or delete items. Every insert or delete requires expanding or shrinking the entire solid-chunk-of-memory allocation, which isn't the fastest thing in the world when your allocation is large. Also, inserting into the HEAD requires copying the _entire_ ziplist up exactly one element position because the block of your memory layout just changed.

So, obviously these things are useful, but their usefulness degrades as your solid blocks of memory hit cache limits. Everything is super fast when your entire ziplist fits in L1, still about the same in L2, worse in L3, and horrible if your ziplist grows any larger.

But, what can we do? We can fix the entire problem by creating a traditional linked lists of solid memory block ziplists. Now, each ziplist can remain small (8 KB to 16 KB), and when we grow bigger, we just cut a new ziplist, pointer-attach it to the previous ziplist, and now we have a minimal-pointer, high-locality data structure with unlimited growth potential without performance defects. It doesn't buckle under inserting or updating arbitrary elements since each memory block is isolated to a maximum ~8 KB size (reminder: your L1 cache is 32 KB to 64 KB depending on architecture) and now your memory usage due to pointer overhead is now also _greatly_ reduced (assuming your data was small and dominated or matched by pointer overhead size in the first place).

As a double bonus, now when your data structure is a solid block of memory, you can compress individual contiguous memory blocks (usually with great compression ratios) because your contents will tend to be homogeneous in shape inside the same container. Result: huge reduction in pointer overhead + sequential access + compression = best of all possible worlds.

Other details/comparisons at https://matt.sh/redis-quicklist-visions

This isn't news. Arrays are fast until you need to insert/delete in them, then they get horrible. This is why we have B-trees. Every page in a B-tree is essentially a sorted array - you can find items in it quickly using binary search, and you can insert/delete in it relatively quickly because it has a small bounded size. When you need to grow beyond a single page you do a page-split and add a parent above it, etc. B-trees are the optimal implementation of dynamic arrays, end of story.

This reply is golden. It starts off completely dismissive with "this isn't news" and it ends completely dismissive with "end of story." I am quite sorry my free internet post failed to meet your quality guidelines.

A wise man once said: Things have now gotten to the stage where I flinch slightly as I click on the "comments" link, bracing myself for the dismissive comment I know will be waiting for me at the top of the page.

For what it is worth I enjoyed your parent post.

If you honestly believe your linked list of blocks is actually superior in efficiency to a B-tree, you're delusional.

>B-trees are the optimal implementation of dynamic arrays

Then why don't the popular scripting languages use B-trees for their sequence containers?

Both Perl and Python use arrays for this. Both pay the price of O(n) inserts mid-container.

I would be unsurprised if it was for legacy reasons. Both of those programming languages were written over twenty years ago. Furthermore, arrays play nicely with C extensions.

Hmmm, let me shoot a quick guess: because B-trees are not an implementation of dynamic arrays?

I think there are several reasons:

1. Name array strongly implies that it has an underlying continuous memory location associated with it. All of modern scripting languages have some form of dictionary, map, set, or hashtable which is implemented using insert/update friendly approach.

1a. Typical workload for array is to iterate over it doing something with each element, B-Tree is strictly worse in iterating over all elements compared to static array.

2. Most instantiated dynamic arrays rarely perform deletes. Most common operation that could change size of array is push_back, and amortized cost for inserts is probably way less than chasing pointers in B-Tree.

3. B-Trees were created and optimized for HDD performance. With current SSD and memory trends, there are simply better data structures to use.

4. Pointer chasing got a lot more expensive relative to continuous memory scanning. This is mostly due to majority of memory performance increases coming from increased throughput as opposed to lower latency.

This does not mean B-Trees are bad, in fact they are still heavily utilized in databases because they scale extraordinarily well with data size. They are commonly used for indexes because membership testing (single value lookup) and range lookups are O(1), and updates/deletes are generally very fast as well.

I think this is a really good answer. However I slightly differ on #1.

Python calls the default sequence container a list. If you took that literally, it implies O(1) inserts.

Interesting question. For some discussion of (among other ADTs) "actual" trees in python, see:

http://kmike.ru/python-data-structures/

And:

http://pythonhosted.org/BTrees/

Cocoa switches away from straight arrays for large "array" objects:

http://ridiculousfish.com/blog/posts/array.html

B-trees are “optimal” in the sense that they’re O(log N) for all the usual operations (insertion/deletion, appenditem, getitem, setitem, and even concatenation with some work) but simple exponentially-growing arrays are O(1) for getitem, setitem, and appenditem, which is asymptotically faster. Also, their constant factors are better, and their memory footprint is smaller.

hyc_symas’s comment seems a little confused because they are talking about doing binary search within B-tree pages. But if you’re using B-trees to implement dynamic non-sparse arrays, which seems to be the topic, you don’t need any binary searches. You can just subtract the beginning-of-page index from the index you’re indexing to in order to get the offset within the page.

That kind of comment, managing to be simultaneously ignorant and arrogantly dismissive, is why I don’t comment much on HN these days.

Sorry, but the only ignorance here is your own. You assumed that "storing into an array" means that the array index is the only thing relevant for locating a record. But the fact that you're talking about inserts/deletes means that the array index of a particular element is mutable and therefore insufficient for uniquely locating a record. Which means you must perform some type of search to retrieve a particular item. Your O(1) getitem is nonsense.

Yes, you can say "give me the 5th element of this array" but the discussion was about maps/dictionaries, where you're more often going to say "give me item foo that was inserted before".

When folks like you post while obviously not paying attention, yes, you elicit a dismissive response.

So if you have fixed sized 'ziplists', why don't you do a memmove() instead of a realloc() on inserts, then split your array on overflow? That would be 1 malloc per 1024 inserts assuming 8 byte data in an 8K page instead of 1 malloc per insert.

This would be vastly more cache friendly, and you would be fragmenting your heap dramatically slower. You could even mmap() a scheme like this, but that's another story!

why don't you do a memmove() instead of a realloc() on inserts

Oh, it's actually worse than that because when inserting we do a realloc (which could require automatically copying the entire memory block), then (if inserting to the head) we memmove the entire contents down to the end of the new allocation. Fun!

Mainly we don't pre-allocate because it's built on top of existing components, and the existing components don't do that. The actual "max size" limit is configurable and we don't really want to allocate the full max size for each list up front. If you have 500,000 lists each with 3 40 byte elements (~60 MB total), it would be overkill to allocate 500,000 8 KB blocks (~4 GB total).

Ideally, we would automatically determine if the user is doing "big operations" then switch to a page-based approach versus smaller "store as much data as compactly as possible" approaches, but it's just not done yet.

Adam Rosenberg wrote a really interesting book advocating not just this but the entire Fortran programming style. He hasn’t persuaded me that this is a good idea in general, but it’s a really interesting way of looking at things: http://www.the-adam.com/adam/rantrave/st02.pdf. If you struggle to imagine how you would structure software without structs or objects, you should definitely read this book!

Among other things, it incidentally solves the problem Yossi Kreinin identified in "I want a struct linker": http://yosefk.com/blog/i-want-a-struct-linker.html

Indeed, this argument is speaks to a fundamental incompatibility between OOP and efficient code. I wonder how to write a compiler to perform this transformation.

Same tension occurs in Java games for phones, at least pre-android. Instead of an array of Sprite objects, turns out its more efficient to have parallel arrays of ints, which leads to less readable code.

with Java it's even worse because SoA is the only way to get your data packed in contiguous memory, as opposed to spraying it all over there GC heap.

The Perl folks call this "inside-out objects"

I guess the MATLAB folks call this "Array Programming"

This is my favorite example of where object oriented programming (and in particular C++) is an advantage.

OO as it is usually practiced (gross generalization I know) is often hell on memory layout and memory access which causes problems that never show up in unit testing but show up in deployment. Pattern fetishism exacerbates the problem.

But a data structure abstraction that hides the storage implementation is an ideal application of OO.

Moderation in all things....

Interestingly Intel's ISPC compiler actually uses a hybrid of the two layouts to store its memory of arrays. I think it stores an array of vectors as xxxx, yyyy, zzzz, wwww, xxxx, ...

1.by Picasa you mean google-picasa?

2.you are saying it's good to use in-memory column-store for oltp?

> "1.by Picasa you mean google-picasa?"

Yes, that Picasa.

Yes, that herf.

Game developers call that principle "Data Oriented Design".

Reply on news.ycombinator.com