Btrees are the new black

http://www.me.net.nz/blog/btrees-are-the-new-black/

by jjh42 • 12 years ago

133 41 12 years ago

me.net.nz

vinkelhake• 12 years ago

Google has an open source implementation of b-trees in C++. You gain performance and memory efficiency (less overhead per element stored). You lose the iterator stability that std::map/std::set guarantees. Other than that the interfaces are pretty much equal.

https://code.google.com/p/cpp-btree/

jjh425• 12 years ago

(top article author). I wasn't aware of that, it looks like they come to similar conclusions "B-trees are widely known as data structures for secondary storage, because they keep disk seeks to a minimum. For an in-memory data structure, the same property yields a performance boost by keeping cache-line misses to a minimum. C++ B-tree containers make better use of the cache by performing multiple key-comparisons per node when searching the tree. Although B-tree algorithms are more complex, compared with the Red-Black tree algorithms, the improvement in cache behavior may account for a significant speedup in accessing large containers."

m_eiman• 12 years ago

A sidenote: your blog post is from the future - unless we're already in May? :)

thesz• 12 years ago

BTrees are cache-aware, and, in a sense, cache-oblivious.

There are more interesting structures, like fractal tree indices, which are cache-oblivios in the very formal sense.

My own opinion is that Btrees are nice, but not very effective in current world. Structures like LSM trees can be more effective. And you can construct high efficiency search structures over runs of LSM tree which will mimic Btrees.

alecco• 12 years ago

Funny how everything Google does is dogma. They move from complicated hash DS to B+trees, a decade later, and everyone follows. Better DS are currently Log-merge for heavy writes and data parallel BSTs for read-most.

At least they are not stubborn for long, I give them that.

yongjik• 12 years ago

I don't know how you got that impression, but inside Google people use hash tables far, far more often than B+trees. In fact, I don't think I ever used B+tree once in my entire career inside Google.

It's pretty hard to beat "expected O(1)".

joe_the_user• 12 years ago

Well,

If the hype is to be believed, Cache-oblivious b-trees are an even better match for today's hierarchical memory systems.

Unfortunately, I don't know any simple free/open-source implementation of these.

But here is a "home page" for them: http://supertech.csail.mit.edu/cacheObliviousBTree.html

JulienSchmidt• 12 years ago

There are even better fits for modern CPU architectures and hierarchical memory systems (but not on HDDs) than B-Tree based structures, like the adaptive radix tree: http://wwwkemper.informatik.tu-muenchen.de/~leis/papers/ART....

I think B-Trees are mostly important for educational purposes, since this is a very important general way of organizing data. For real world usage, there are hundreds of different structures optimized for different use cases.

justin66• 12 years ago

Do you know of any examples of good open source projects using the ART?

checker659• 12 years ago

What about false sharing?

jjh425• 12 years ago

I only benchmarked and considered the single-threaded case. Multithreading is a whole new can of worms.

warmfuzzykitten2• 12 years ago

Trivially, one could synchronize around the entire tree.

However, if modified nodes are always copied, then the only node that matters is the root node (since there can never be a partially modified tree accessible from the root), which boils down to atomic update of a root node pointer. Should be very efficient.

There is, of course, a performance implication of node copying, but it only affects add/delete/replace performance - read access runs at full speed - and it is cache-friendly, as well.

hyc_symas2• 12 years ago

LMDB was written specifically for multi-threading; all of its shared data structures are cache-line aligned to prevent false sharing. This is also essential to allowing readers to run with no locks.

cheepin• 12 years ago

Interestingly, a red-black tree can be viewed as a BTree [0].

Another case where caching plays a huge role in determining the most efficient data structure (See vector vs list [1]).

[0] http://en.wikipedia.org/wiki/Red%E2%80%93black_tree#Analogy_...

[1] http://baptiste-wicht.com/posts/2012/12/cpp-benchmark-vector...

nonsens3• 12 years ago

Robert Sedgewick gives a very good analogy of a red-black tree to a 2-3 tree on his Coursera course (Algorithms, Part I) [0] - a bit different to what is discussed in Wikipedia, but very easy to understand and implement.

[0] https://class.coursera.org/algs4partI-004

csch• 12 years ago

Timo Bingmann presented a fast and robust implementation here https://panthema.net/2007/stx-btree/

arneeiri• 12 years ago

This reminds me of The Ubiquitous B-Tree by Douglas Comer. That article was written in 1978.

http://people.cs.aau.dk/~simas/aalg06/UbiquitBtree.pdf

beagle3• 12 years ago

The original B-Trees as described by Bayer were "intrusive" (like implemented here), in the sense that every node contains keys, values, and pointers.

There are variants such as the B+Tree http://en.wikipedia.org/wiki/B%2B_tree , which stores only keys in nodes and chains blocks - which is more efficient in range scans and less in general retrieval; And the http://en.wikipedia.org/wiki/B*-tree which is more densely packed.

srean2• 12 years ago

Since there has been a resurgence of interest in OCaml on HN, some may be curious to see how a production quality Btree implementation looks like in OCaml. Algebraic data types, make it particularly convenient for these sorts of things. For comparison one may contrast it with an implementation in a language that does not have algebraic types, C++ for example. So glad Rust chose to have them.

Here is a Btree implementation from the mirage project [1] https://github.com/cgreenhalgh/ocaml-btree/blob/master/lib/b...

EDIT: removed link to binary tree after kerneis pointed it out.

[1] http://www.openmirage.org/

narag• 12 years ago

Sweet. I remember writing an implementation in TurboPascal in the nineties. This is... refreshing :-)

kerneis2• 12 years ago

It can be shorter still, if all you want is insert and read

Despite the name, I'm afraid this is a simple binary tree implementation, not a BTree:

type 'a tree = Empty | Node of('a * 'a tree * 'a tree);;

Obviously the author didn't know what he was doing (the intended scope was very large, AVL, BTrees, etc. but he stopped after committing this single file).

kerneis2• 12 years ago

Here is a Btree implementation from the mirage project

And in fact, this implementation does not show a lot because the actual B-tree work is done by the baardskeerder library: https://github.com/Incubaid/baardskeerder

It looks very large and complex, it is most certainly possible to write a simpler, pedagogical implementation.

srean2• 12 years ago

Found cpc https://github.com/kerneis/cpc from your comment stream. Looks super interesting. http://felix-lang.org/share/src/web/tut/fibres_index.fdoc might pique your interest. It also compiles coroutines portably and efficiently (to a C++ runtime library). Felix tries to do for C++ what F# does to C# (in case there is any confusion, I am just a new user)

ef47d35620c1• 12 years ago

In C++, std::set and std::map are typically implemented as RB trees. To compare the performance of those to a hash backed container, try std::unordered_set and std::unordered_map.

bdg• 12 years ago

Someone should create a library where all your data is stored in a BTree. Think of it! Huge data sets, small data sets, everything scales in this system. And even better, they can create a language-agnostic API to a small server that just stores the btrees for them and gives them back on request. We could call it a database, but that would require we use indexes correctly for once in our lives.

rplnt• 12 years ago

NTFS uses BTrees I believe. And matches your requirements.

taeric2• 12 years ago

Mayhap I'm the one missing the joke. That being that this is just describing databases, period.

erichocean• 12 years ago

If you do want a B-tree library in C that's not tied to SQL, I've had good success with lmdb[0]. As a bonus, it's also transactional and writers do not block readers (and vice versa). Crazy fast and easier to use from C than SQLite IMO.

If you want B-trees that compete with LSM trees, stratified B-trees[1] are pretty sweet. There's an in-kernel GPL implementation called Castle[2] on GitHub.

[0] http://symas.com/mdb/

[1] http://www.slideshare.net/acunu/20110620-stratifiedbtree

[2] https://github.com/acunu/castle

ebiester2• 12 years ago

And then we can stack every conceivable use case on top of it, and write that language-agnostic API in a way that makes interfacing with most current languages so onerous that people can build hundreds of thousands of lines of code to make it slightly less painful!

Maybe we can make the configuration so convoluted that we can base entire companies and revenue models around maintaining them.

coldtea• 12 years ago

write that language-agnostic API in a way that makes interfacing with most current languages so onerous that people can build hundreds of thousands of lines of code to make it slightly less painful

Those would be people who couldn't be bothered to learn a really simple declarative language, and prefer ad-hoc OO ORMS and other shit like that...

briandw• 12 years ago

Under what conditions will a B-tree out perform a hash map? Seems like the hash map was far and away the better choice in his test.

jjh425• 12 years ago

For insertions and lookup it is probably hard to find a situation (for large n hashes are asymptotically superior).

Trees provide some other nice properties such as maintaining an ordering (so it's fast to get the smallest argument or interate of the elements in numerical order) and always perform at the asymptotic time cost (a dynamic hash table occasionally takes O(n) when the table is resized so it might not be a good fit in a realtime situation)

anotherevan• 12 years ago

I have fond memories of red/black binary trees. About 15 years ago I wrote a reusable library for R/B trees in C. I had a series of programs I had to write that were collating a lot of data in memory into several buckets.

(The nice thing about virtual memory is you can make a really big RAM disk.)

thezorg• 12 years ago

Speaking of B-trees, I find CouchDB's implementation really interesting: the backing file is append-only, meaning that modifying the tree implies appending the modified nodes at the end of the file (for example, if I change a leaf, than I rewrite it at the end of the file, and rewrite its parent so that it points to the leaf's new position, and so on until we reach the root node).

Sounds like a waste of space, but it solves a bunch of problems, like being crash-proof (since we never overwrite anything, it's always possible to just find the last root) and allowing reads concurrent with a write without any synchronizing necessary. Plus, it's actually optimized for SSD's due to its log-like structure!

CouchDB B-Trees : http://guide.couchdb.org/draft/btree.html Log structures in SSDs : http://lwn.net/Articles/353411/

hyc_symas2• 12 years ago

Massive waste of space though. LMDB is also copy-on-write and crash-proof but doesn't require compaction like CouchDB does.

http://symas.com/mdb/

Read the LDAPCon 2011 paper (and slides) to see how it works and why it's better than CouchDB.

Btrees are the new black

Related Stories