This repo contains a Machine-Learning enhanced B+ Tree index. You can call it learned B+ Tree or Model-based B+ Tree. It replaces the data format of internal nodes with a linear regression model and a gapped array.
For each internal node, we first train a linear model on the <key, pointer> pairs of data inside the node, then reoganize the positions of the pairs based on the model prediction.
The benefit is two-fold:
- The binary search in each internal node is replaced by a prediction-based exponential search.
- This does not consume more memory space compared to the original B+ Tree implementation. In most cases, a B+ Tree node is only around 50% - 70% full for the sake of insertion.
When a B+ node is close to full, the prediction-based search might be slightly worse than binary search. Because the accuracy of model prediction will downgrade while the performance of exponential search is highly affected by the starting position of the search.
This implementation supports bulk-load and insertion.
Please read the test cases: https://github.com/jiayuasu/stx-btree/blob/master/testsuite/BulkLoadTest.cc
The performance of this index is studied in our SIGMOD 2020 paper:
Ding, Jialin, Umar Farooq Minhas, Jia Yu, Chi Wang, Jaeyoung Do, Yinan Li, Hantian Zhang et al. "ALEX: an updatable adaptive learned index." In Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data, pp. 969-984. 2020.
A PDF copy of this paper is available on ACM website: https://dl.acm.org/doi/pdf/10.1145/3318464.3389711
Figure 9 illustrates the performance of this Model B+ Tree:
Author: Timo Bingmann (Mail: tb a-with-circle panthema dot net)
Date: 2013-05-05
https://github.com/tlx/tlx/tree/master/tlx/container
This is an improved version with better STL semantics and it will be maintain in the TLX library.
The STX B+ Tree package is a set of C++ template classes implementing a B+ tree key/data container in main memory. The classes are designed as drop-in replacements of the STL containers set, map, multiset and multimap and follow their interfaces very closely. By packing multiple value pairs into each node of the tree the B+ tree reduces heap fragmentation and utilizes cache-line effects better than the standard red-black binary tree. The tree algorithms are based on the implementation in Cormen, Leiserson and Rivest's Introduction into Algorithms, Jan Jannink's paper and other algorithm resources. The classes contain extensive assertion and verification mechanisms to ensure the implementation's correctness by testing the tree invariants. To illustrate the B+ tree's structure a wxWidgets demo program is included in the source package.
The current source package can be downloaded from http://panthema.net/2007/stx-btree/
The include files are extensively documented using doxygen. The compiled doxygen html documentation is included in the source package. It can also be viewed online at http://panthema.net/2007/stx-btree/stx-btree-0.9/doxygen-html/
The wxWidgets B+ tree demo program is located in the directory wxbtreedemo. Compiled binary versions can be found on the package web page.
If bugs should become known they will be posted on the above web page together with patches or corrected versions.
The B+ tree template source code is released under the Boost Software License, Version 1.0, which can be found at the header of each include file.
All auxiliary programs like the wxWidgets demo, test suite and speed tests are licensed under the GNU General Public License v3 (GPLv3), which can be found in the file COPYING.GPLv3.
The idea originally arose while coding a read-only database, which used a huge map of millions of non-sequential integer keys to 8-byte file offsets. When using the standard STL red-black tree implementation this would yield millions of 20-byte heap allocations and very slow search times due to the tree's height. So the original intension was to reduce memory fragmentation and improve search times. The B+ tree solves this by packing multiple data pairs into one node with a large number of descendant nodes.
In computer science lectures it is often stated that using consecutive bytes in memory would be more cache-efficient, because the CPU's cache levels always fetch larger blocks from main memory. So it would be best to store the keys of a node in one continuous array. This way the inner scanning loop would be accelerated by benefiting from cache effects and pipelining speed-ups. Thus the cost of scanning for a matching key would be lower than in a red-black tree, even though the number of key comparisons are theoretically larger. This second aspect aroused my academic interest and resulted in the speed test experiments.
A third inspiration was that no working C++ template implementation of a B+ tree could be found on the Internet. Now this one can be found.
This implementation contains five main classes within the stx namespace (blandly named Some Template eXtensions). The base class btree implements the B+ tree algorithms using inner and leaf nodes in main memory. Almost all STL-required function calls are implemented (see below for the exceptions). The asymptotic time requirements of the STL standard are theoretically not always fulfilled. However in practice this B+ tree performs better than the STL's red-black tree at the cost of using more memory. See the speed test results for details.
The base class is then specialized into btree_set, btree_multiset, btree_map and btree_multimap using default template parameters and facade functions. These classes are designed to be drop-in replacements for the corresponding STL containers.
The insertion function splits the nodes on recursion unroll. Erase is largely based on Jannink's ideas. See http://dbpubs.stanford.edu:8090/pub/1995-19 for his paper on "Implementing Deletion in B+-trees".
The two set classes (btree_set and btree_multiset) are derived from the base implementation class btree by specifying an empty struct as data_type. All functions are adapted to provide the base class with empty placeholder objects. Note that it is somewhat inefficient to implement a set or multiset using a B+ tree: a plain B tree (without +) would hold no extra copies of the keys. The main focus was on implementing the maps.
The most noteworthy difference to the default red-black tree implementation of std::map is that the B+ tree does not hold key/data pairs together in memory. Instead each B+ tree node has two separate arrays containing keys and data values. This design was chosen to utilize cache-line effects while scanning the key array.
However it also directly generates many problems in implementing the iterators' operators. These return a (writable) reference or pointer to a value_type, which is a std::pair composition. These data/key pairs however are not stored together and thus a temporary copy must be constructed. This copy should not be written to, because it is not stored back into the B+ tree. This effectively prohibits use of many STL algorithms which writing to the B+ tree's iterators. I would be grateful for hints on how to resolve this problem without folding the key and data arrays.
The B+ tree distribution contains an extensive test suite. According to gcov 90.9% of the btree.h implementation is covered.
The tree algorithms currently do not use copy-construction. All key/data items are allocated in the nodes using the default-constructor and are subsequently only assigned new data (using operator=).
The most important incompatibility are the non-writable operator* and operator-> of the iterator. See above for a discussion of the problem on separated key/data arrays. Instead of *iter and iter-> use the new function iter.data() which returns a writable reference to the data value in the tree.
The B+ tree supports three erase functions:
size_type erase(const key_type &key); // erase all data pairs matching key bool erase_one(const key_type &key); // erase one data pair matching key void erase(iterator iter); // erase pair referenced by iter
The following STL-required function is not supported:
void erase(iterator first, iterator last);
Beyond the usual STL interface the B+ tree classes support some extra goodies.
// Output the tree in a pseudo-hierarchical text dump to std::cout. This
// function requires that BTREE_DEBUG is defined prior to including the btree
// headers. Furthermore the key and data types must be std::ostream printable.
void print() const;
// Run extensive checks of the tree invariants. If a corruption in found the
// program will abort via assert(). See below on enabling auto-verification.
void verify() const;
// Serialize and restore the B+ tree nodes and data into/from a binary image.
// This requires that the key and data types are integral and contain no
// outside pointers or references.
void dump(std::ostream &os) const;
bool restore(std::istream &is);
All tree template classes take a template parameter structure which holds important options of the implementation. The following structure shows which static variables specify the options and the corresponding defaults:
struct btree_default_map_traits
{
// If true, the tree will self verify it's invariants after each insert()
// or erase(). The header must have been compiled with BTREE_DEBUG
// defined.
static const bool selfverify = false;
// If true, the tree will print out debug information and a tree dump
// during insert() or erase() operation. The header must have been
// compiled with BTREE_DEBUG defined and key_type must be std::ostream
// printable.
static const bool debug = false;
// Number of slots in each leaf of the tree. Estimated so that each node
// has a size of about 128 bytes.
static const int leafslots =
MAX( 8, 128 / (sizeof(_Key) + sizeof(_Data)) );
// Number of slots in each inner node of the tree. Estimated so that each
// node has a size of about 128 bytes.
static const int innerslots =
MAX( 8, 128 / (sizeof(_Key) + sizeof(void*)) );
// As of stx-btree-0.9, the code does linear search in find_lower() and
// find_upper() instead of binary_search, unless the node size is larger
// than this threshold. See notes at
// http://panthema.net/2013/0504-STX-B+Tree-Binary-vs-Linear-Search
static const size_t binsearch_threshold = 256;
};
The implementation was tested using the speed test sources contained in the package. For a long discussion on results please see the web page