PowerPoint Presentation
EECS 4101/5101
B-trees
2-3-4 trees
Prof. Andy Mirzaian
Lists
Move-to-Front
Search Trees
Binary Search Trees
Multi-Way Search Trees
B-trees
Splay Trees
2-3-4 Trees
Red-Black Trees
SELF ADJUSTING
WORST-CASE EFFICIENT
competitive
competitive?
Linear Lists
Multi-Lists
Hash Tables
DICTIONARIES
2
References:
[CLRS] chapter 18
3
B-trees
R. Bayer, E.M. McCreight,
“Organization and maintenance of large ordered indexes,”
Acta Informatica 1(3), 173-189, 1972.
Boeing Company
B…
4
Definition
B-trees are a special class of multi-way search trees.
Node size:
d[x] = degree of node x, i.e., number of subtrees of x.
d[x] – 1 = number of keys stored in node x. (This is n[x] in [CLRS].)
Definition: Suppose d 2 is a fixed integer.
T is a B-tree of order d if it satisfies the following properties:
1. Search property: T is a multi-way search tree.
2. Perfect balance: All external nodes of T have the same depth (h).
3. Node size range: d d[x] 2d nodes x root[T]
2 d[x] 2d for x = root[T].
5
Example
36
12 18 26
42 48
2 4 6 8 10
14 16
20 22 24
28 30 32 34
38 40
44 46
50 52 54
h = 3
height,
including
external
nodes
A B-tree of order d = 3 with n = 27 keys.
6
Warning
Perfect Balance: All external nodes of T have the same depth.
The only leaf in this tree
[CLRS] says: All leaves have the same depth.
7
Applications
External Memory Dictionary:
Large dictionaries stored in external memory.
Each node occupies a memory page.
Each node access triggers a slow page I/O.
Keep height low. Make d large.
A memory page should hold the largest possible node (of degree 2d).
Typical d is in the range 50 .. 2000 depending on page size.
Internal Memory Dictionary:
2-3-4 tree = B-tree of order 2.
8
B-tree Height
How small or large can a B-tree of order and height be?
keys in the B-tree
external nodes, all at depth
This is lower/upper bounded by the min/max aggregate branching
at depths :
Take logarithm:
9
Height in B-trees & 2-3-4 trees
B-trees:
2-3-4 trees:
This includes the level of external nodes.
10
SEARCH
Algorithm:
Simply follow the search path for the given key.
Complexity:
At most O(h) nodes probed along the search path.
Each node has O(d) keys.
# I/O operations:
O(h) = O( logd n) = O(log n / log d).
So, for external memory use keep d high.
Page size is the limiting factor.
Search time:
binary search probing within node: O(h log d) = O(log n).
sequential probing within node: O(hd) = O(d log n / log d).
INSERT & DELETE (to be shown) also take O(hd) time.
So, for internal memory use keep d low (e.g., 2-3-4 tree).
11
Local Restructuring
INSERT and DELETE need the following local operations:
Node splitting
Node fusing
Key sharing (or key borrowing)
The first one is used by INSERT.
The last 2 are used by DELETE.
Each of them takes O(d) time.
12
Node Splitting & Fusing
Split(x)
Parent p gains a new key. (If p was full, it needs repair too.)
If x was the root, p becomes the new root (of degree 2) & tree height grows by 1.
Fuse (x’, ad, x’’)
Parent p loses a key. (If non-root p was a d-node, it needs repair too.)
If p was the root & becomes empty, x becomes the new root & tree height shrinks by 1.
O(d)
time
p
… bi-1 bi …
T1
Td
Td+1
T2d
a1 … ad-1 ad ad+1 … a2d-1
x
…
…
a
… bi-1 ad bi …
T1
Td
x’
p
Td+1
T2d
……
…….
a1 … … ad-1
ad+1 … a2d-1
x”
a’
a”
13
Key Sharing
KeyShare(x,y)
Node y is the (left or right) immediate sibling of x.
d[x]=d, d[y]>d.
x “borrows” a key from y to make d[x] > d.
Note: the inorder sequence of keys is not disturbed.
O(d)
time
… ci-1 ci ci+1 …
T1
Td
x
p
Td+1
Td+2
……
…
a1 … … ad-1
b1 b2 b3 …
y
Td+3
… ci-1 b1 ci+1 …
T1
Td
x
p
Td+1
Td+2
……
…
a1 … … ad-1 ci
b2 b3 …
y
Td+3
14
Bottom-Up INSERT
Bottom-Up INSERT (K,T)
Follow the search path of K down B-tree T.
if K is found then return
Otherwise, we are at a leaf x where K needs to be inserted.
K’ K (* the insertion key at the current node x *)
while d[x] = 2d do (* repeated node splitting up the search path *)
Split x into x’ and x’’ with middle key keyd[x]
if K’ < keyd[x] then insert K’ into x’ else insert K’ into x”
K’ keyd[x]
x p[x]
end-while
if x nil then insert K’ into x
else do
create a new node r
root[T] r
insert K’ into r
make x’ and x’’ the two children of r
end-else
end
O(d logd n) time
15
Top-Down INSERT
Top-Down INSERT (K,T)
As you follow the search path of K down B-tree T, keep
splitting every full node along the way and insert its middle
key into its (now non-full) parent.
If the root was full and got split, a new degree 2 root would
be created during this process.
if K is found then return
else insert K into the (now non-full) leaf you end up.
end
Pros and Cons:
Top-Down Insert tends to over split. It could potentially split O(h) nodes,
while the Bottom-Up Insert may not split at all!
Top-Down makes only one pass down the search path and does not need
parent pointers, while Bottom-Up may need to climb up to repair by repeated
node splitting.
O(d logd n) time
16
Top-Down DELETE
Top-Down DELETE (K,T)
The idea is to move the current node x down the search path of K in T and maintain: Loop Invariant: d[x] > d or x=root[T].
By LI, x can afford to lose a key.
At the start x = root[T] and LI is maintained.
Case 0: Kx and x is a leaf. Return
Case 1: Kx and we have to move to child y of x.
Case 1a: d[y] > d:
x y (*LI is maintained *)
Case 1b: d[y] = d, d[z] > d, (z=left/right imm. sib. of y)
KeyShare(y,z);
x y (*LI is maintained *)
Case 1c: d[y] = d & d[z]=d separated by key D at x:
x Fuse(y,D,z) (* LI is maintained *)
Case 2: Kx and we have to remove K from x:
see next page
Continued
x
root[T]
… D …
y
z
x
17
Top-Down DELETE
Top-Down DELETE (K,T)
Case 2: Kx and we have to remove K from x:
x
root[T]
… K …
y
z
x
K’
K”
Case 2a: x is a leaf: (* LI *)
remove K from x
If x=root[T] & becomes empty, set root[T] nil
Case 2b: x is not a leaf:
Let y and z be children of x imm. left/right of K
Case 2b1: d[y] > d (* LI *)
Recursively remove predecessor K’ of K from
subtree rooted at y and replace K in x by K’.
Case 2b2: d[y] = d, d[z] > d (* LI *)
Recursively remove successor K’’ of K from
subtree rooted at z and replace K in x by K”.
Case 2b3: d[y] = d[z] = d
x Fuze(y,K,z) (* LI *)
Repeat Case 2 (one level lower in T).
O(d logd n) time
Bottom-Up
Delete
left as exercise
18
2-3-4 trees
19
2-3-4 tree
This includes the level of external nodes.
36
12 18 26
42 48
6 8 10
14 16
20 22 24
28 30 32
38 40
44 46
50 52 54
h = 3
height,
including
external
nodes
2-3-4 tree is a B-tree of order 2, i.e., a multi-way search tree with
2-nodes, 3-nodes, 4-nodes, and all external nodes at the same depth.
20
Example: Bottom-Up Insert
h=4
c, d, e
j, k, l
a, b, m
A
B
C
D
E
F
f, h, i
G
H
I
J
K
L
M
Insert g
A
B
C
D
E
F
G’
H
I
J
K
L
M
G”
i
f, g
l
h, j
e, k
c
d, m
a
b
h=5
Insert g
21
Exercises
22
Bottom-Up Delete: Design and analyze an efficient implementation of bottom-up Delete on B-trees.
Insert & Delete Examples: Demonstrate the following processes, each time starting with the B-tree of order 3 example on page 6 of these slides.
(a) Insert 1.
(b) Delete 48 top-down.
(c) Delete 48 bottom-up.
B-tree Insertion sequence: Draw the B-tree of order d resulting from inserting the following keys, in the given order, using bottom-up insert, into an initially empty tree:
4, 40, 23, 50, 11, 34, 62, 78, 66, 22, 90, 59, 25, 72, 64, 77, 39, 12,
(a) with d = 3.
(b) with d = 4.
2-3-4 tree Insertion sequence: Insert integers 1..32 in that order into an initially empty 2-3-4 tree.
(a) Show some intermediate snapshots as well as the final 2-3-4 tree.
(b) Would you get a different resulting tree with top-down versus bottom-up Insert?
(c) Can you describe the general pattern for the insertion sequence 1..n?
[Hint: any connection with the Binary Counter with the Increment operation?]
Split and Join on 2-3-4 trees: These are cut and paste operations on dictionaries. The Split operation takes as input a dictionary (a set of keys) A and a key value K (not necessarily in A), and splits A into two disjoint dictionaries B = { xA | key[x] K } and C = { xA | key[x] > K }. (Dictionary A is destroyed as a result of this operation.) The Join operation is essentially the reverse; it takes two input dictionaries A and B such that every key in A < every key in B, and replaces them with their union dictionary C = AB. (A and B are destroyed as a result of this operation.) Design and analyze efficient Split and Join on 2-3-4 trees. [Note: Definition of Split and Join here is the same we gave on BST’s and slightly different than the one in [CLRS, Problem 18-2, pp: 503-504].]
23
B*-trees: A B*-tree T (of order d>0) is a variant of a B-tree. The only difference is the node size range. More specifically, for each non-root node x, 2d d[x] 3d
( i.e., at least 2/3 full. )
(a) Specify appropriate lower and upper bounds on d[root[T]] of a B*-tree.
(b) Describe an insertion procedure for B*-trees. What is the running time?
[Hint: Before splitting a node see whether its sibling is full. Avoid splitting if possible.)
(c) What are the advantages of a B*-tree over a standard B-tree?
24
END
25
.
2
1
n
d
d
2
log
1
h
)
1
n
(
log
+
+
£
£
+
.
)
1
n
(
log
h
)
1
n
(
log
2
2
2
1
+
£
£
+
/docProps/thumbnail.jpeg