Download presentation

Presentation is loading. Please wait.

Published byAnahi Rathbun Modified about 1 year ago

1
Multidimensional Data Rtrees Bitmap indexes

2
R-Trees For “regions” (typically rectangles) but can represent points. Supports NN, “whereamI” queries. Generalizes Btree to multidimensional case. - In place of Btree's keypointer pairs, Rtree has region pointer pairs.

3
Lookup – Where Am I? We start at the root, with which the entire region is associated. We examine the subregions at the root and determine which children correspond to interior regions that may contain point P. If there are zero regions we are done; P is not in any data region. If there are some subregions we must recursively search those children as well, until we reach the leaves of the tree.

4
Insertion Inserting region R. 1. We start at the root and try to find some subregion into which R fits. –If more than one we pick just one, and repeat the process there. 2. If there is no region, we expand, and we want to expand as little as possible. –So, we pick the child that will be expanded as little as possible. 3. Eventually we reach a leaf, where we insert region R. 4. However, if there is no room we have to split the leaf. –We split the leaf in such a way as to have the “smallest subregions.”

5
Example Suppose that the leaves have room for six regions. Further suppose that the six regions are together on one leaf, whose region is represented by the outer solid rectangle. Now suppose that another region POP is added.

6
((0,0),(60,50))((20,20),(100,80)) Road1 Road2 House1School House2 Pipeline Pop Example (Cont’ ed)

7
Suppose now that House3 ((70,5),(80,15)) gets added. We do have space in the leaves, but we need to expand one of the regions at the parent. We choose to expand the one which needs to be expanded the least.

8
Which one should we expand? ((0,0),(80,50))((20,20),(100,80)) Road1 Road2 House1 House3School House2 Pipeline Pop ((0,0),(60,50))((20,5),(100,80)) Road1 Road2 House1School House2 Pipeline Pop House3 Two choices: We chose this option because it produces the smallest increase in total surface area

9
Bitmap Indexes

10
Suppose we have n tuples. A bitmap index for a field F is a collection of bit vectors of length n, one for each possible value that may appear in the field F. The vector for value v has 1 in position i if the i-th record has v in field F, and it has 0 there if not. (30, foo) (30, bar) (40, baz) (50, foo) (40, bar) (30, baz) foo bar 0… baz …

11
Graphical Picture MF CustidNameGenderRating 112JoeM3 115SamM5 119SueF5 112WuM Two bit strings for the Gender bitmap Customer table. We will index Gender and Rating. Note that this is just a partial list of all the records in the table Five bit strings for the Rating bitmap

12
Bitmap operations Bit maps are designed to support partial match and range queries. How? To identify the records holding a subset of the values from a given dimension, we can do a binary OR on the bitmaps from that dimension. - For example, the OR of bit strings for Age = (20, 21, 22) To identify the partial matches on a group of dimensions, we can simply perform a binary AND on the OR-ed maps from each dimension. These operations can be done very quickly since binary operations are natively supported by the CPU.

13
Bit Map example MF SELECT * FROM Customer WHERE gender = M AND (rating = 3 OR rating = 5) AND= First two records in our fact table are retrieved OR=

14
Gold-Jewelry Data (25; 60) (45; 60) (50; 75) (50; 100) (50; 120) (70; 110) (85; 140) (30; 260) (25; 400) (45; 350) (50; 275) (60; 260) What would be the bitmap index for age, and the bitmap index for salary? Suppose we want to find the jewelry buyers with an age in the range and a salary in the range What do we do?

15
How big do these things get? Assuming each attribute value fits in a 32-bit machine word, the bitmap index for an attribute with value cardinality 32 takes as much space as the base data column. Since a B-tree index for a 32-bit attribute is often observed to use 3 or 4 times the space as the base data column, many users consider attributes with cardinalities less than 100 to be suitable for using bitmap indices. However, some other users believe bit map indexes are good for attributes with cardinalities more than 100. What can be done?

16
Basic Compression Run length encoding is used to encode sequences or runs of zeros. Say that we have 20 zeros, then a 1, then 30 more zeros, then another 1. Naively, we could encode this as the integer pair - This would work. But what's the problem? - On a typical 32-bit machine, an integer uses 32 bits of storage. So our pair uses 64 bits. The original string only had 52!

17
Basic Compression (Cont'd) So we must use a technique that stores our run-lengths as compactly as possible. Let’s say we have the string This is made up of runs with 3 zeros and 1 zero. - In binary, 3 = 11, while 1 is, of course, just 1 - This gives us a compressed representation of 111. The problem? - How do we decompress this? - We could interpret this as 1-11 or 11-1 or even This would give us three different strings after the decompression.

18
Proper RLE encoding Run of length i has i 0’s followed by a 1. Let’s say that that we have a run of length i. Let j be the number of bits required to represent i. To define a run, we will use two values: - The “unary” representation of j A sequence of j – 1 “1” bits followed by a zero (the zero signifies the end of the unary string) The special cases of j = 0 and j = 1 use 00 and 01 respectively. - The value i (using j bits) Here we have two “0” runs of length 13 and 6 13 can be represented by 4 bits, 6 requires 3 bits Run 1: j – 1 “1” bits i → Run 2: j – 1 “1” bits i → Final compressed string: Compression rate: (21-14)/21 = 33% Example:

19
Decoding Let’s decode 3 Our sequence of run lengths is: 13, 0, 3. What’s the bitmap?

20
Summary Pros: 1.Bitmaps provide efficient storage for low cardinality dimensions. On sparse, high cardinality dimensions, compression can be effective. 2.Bit operations can support multi-dimensional partial match and range queries Cons: 1.De-compression requires run-time overhead 2.Bit operations on large maps and with large dimension counts can be expensive.

Similar presentations

© 2016 SlidePlayer.com Inc.

All rights reserved.

Ads by Google