U.S. patent application number 12/493210 was filed with the patent office on 2010-12-30 for dictionary-based order-preserving string compression for main memory column stores.
Invention is credited to Carsten Binnig, Franz Faerber, Stefan Hildenbrand.
Application Number | 20100328115 12/493210 |
Document ID | / |
Family ID | 42646299 |
Filed Date | 2010-12-30 |
View All Diagrams
United States Patent
Application |
20100328115 |
Kind Code |
A1 |
Binnig; Carsten ; et
al. |
December 30, 2010 |
DICTIONARY-BASED ORDER-PRESERVING STRING COMPRESSION FOR MAIN
MEMORY COLUMN STORES
Abstract
Methods and systems are described that involve usage of
dictionaries for compressing a large set of variable-length string
values with fixed-length integer keys in column stores. The
dictionary supports updates (e.g., inserts of new string values)
without changing codes for existing values. Furthermore, a
shared-leaves approach is described for indexing such a dictionary
that compresses the dictionary itself while offering access paths
for encoding and decoding.
Inventors: |
Binnig; Carsten; (Elztal,
DE) ; Faerber; Franz; (Walldorf, DE) ;
Hildenbrand; Stefan; (Altdorf, CH) |
Correspondence
Address: |
SAP AG
3410 HILLVIEW AVENUE
PALO ALTO
CA
94304
US
|
Family ID: |
42646299 |
Appl. No.: |
12/493210 |
Filed: |
June 28, 2009 |
Current U.S.
Class: |
341/51 |
Current CPC
Class: |
G06F 16/24561 20190101;
G06F 16/24537 20190101; G06F 16/24557 20190101 |
Class at
Publication: |
341/51 |
International
Class: |
H03M 7/34 20060101
H03M007/34 |
Claims
1. A machine-readable storage medium tangibly storing
machine-readable instructions thereon, which when executed by the
machine, cause the machine to perform operations comprising:
propagating a plurality of string values to a compressed leaf data
of shared-leaves structure of a dictionary via an encode index;
obtaining a plurality of order-preserving integer codes for the
plurality of string values via an lookup operation; if a subset of
the plurality of integer codes was not found during the obtaining,
inserting a subset of the plurality of string values for which the
subset of the plurality of integer codes was not found into the
shared-leaves structure; generating the subset of the plurality of
integer codes for the corresponding subset of the plurality of
string values; and providing a list of the order-preserving
plurality of integer codes, the list including the generated subset
of the plurality of integer codes.
2. The machine-readable storage medium of claim 1 wherein the
operations further comprise: propagating the plurality of integer
codes to the shared-leaves structure of the dictionary via a decode
index; and updating the encode index and the decode index.
3. The machine-readable storage medium of claim 1, wherein the
operations further comprise: rewriting a string value from the
plurality of string values in an equality-predicate or in a
range-predicate with a corresponding integer code from the
plurality of order-preserving integer codes; and rewriting a string
prefix of a prefix-predicate with corresponding integer code ranges
from the plurality of order-preserving integer codes.
4. The machine-readable storage medium of claim 1, wherein
obtaining the plurality of order-preserving integer codes further
comprises executing a sequential search operation over the
compressed leaf data without decompression of the compressed leaf
data.
5. The machine-readable storage medium of claim 1, wherein the
encode index comprise a cache-sensitive array trie index or a
cache-sensitive prefix tree index.
6. The machine-readable storage medium of claim 5, wherein the
cache-sensitive array trie comprises: storing the plurality of
string values in an array; propagating in preorder the plurality of
string values to the shared-leaves structure via variable buffers
at each cache-sensitive array trie node to populate the array only
once per bulk; and generating the subset of the plurality of
integer codes for the corresponding subset of the plurality of
string values in parallel.
7. The machine-readable storage medium of claim 5, wherein the
cache-sensitive prefix tree comprises: calculating a first shortest
prefix to distinguish a largest value of a first leaf and a
smallest value of a second leaf of the cache-sensitive prefix tree;
calculating a second shortest prefix to distinguish the largest
value of the second leaf and the smallest value of a third leaf; if
there is more than one node in a level of the cache-sensitive
prefix tree, adding a second level on top with a node storing the
calculated first and second prefixes, wherein the node is a root of
the cache-sensitive prefix tree.
8. A computer implemented method comprising: propagating a
plurality of string values to compressed leaf data of a
shared-leaves structure of a dictionary via an encode index;
obtaining a plurality of order-preserving integer codes for the
plurality of string values via an lookup operation; if a subset of
the plurality of integer codes was not found during obtaining,
inserting a subset of the plurality of string values for which the
subset of the plurality of integer codes was not found into the
shared-leaves structure; generating the subset of the plurality of
integer codes for the corresponding subset of the plurality of
string values; and providing a list of the order-preserving
plurality of integer codes, the list including the generated subset
of the plurality of integer codes.
9. The method of claim 8 further comprising: propagating the
plurality of integer codes to the shared-leaves structure of the
dictionary via a decode index; and updating the encode index and
the decode index.
10. The method of claim 8 further comprising: rewriting a string
value from the plurality of string values in an equality-predicate
or in a range-predicate with a corresponding integer code from the
plurality of order-preserving integer codes; and rewriting a string
prefix of a prefix-predicate with corresponding integer code ranges
from the plurality of order-preserving integer codes.
11. The method of claim 8, wherein obtaining a plurality of
order-preserving integer codes further comprises executing a
sequential search operation over the compressed leaf data without
decompression of the compressed leaf data.
12. The method of claim 8, wherein the encode index comprises a
cache-sensitive array trie index or a cache-sensitive prefix tree
index.
13. The method of claim 12, wherein the cache-sensitive array trie
comprises: storing the plurality of string values in an array;
propagating in preorder the plurality of string values to the
shared-leaves structure via variable buffers at each
cache-sensitive array trie node to populate the array only once per
bulk; and generating the subset of the plurality of integer codes
for the corresponding subset of the plurality of string values in
parallel.
14. The method of claim 12, wherein the cache-sensitive prefix tree
comprises: calculating a first shortest prefix to distinguish a
largest value of a first leaf and a smallest value of a second leaf
of the cache-sensitive prefix tree; calculating a second shortest
prefix to distinguish the largest value of the second leaf and the
smallest value of a third leaf; if there is more than one node in a
level of the cache-sensitive prefix tree, adding a second level on
top with a node storing the calculated first and second prefixes,
wherein the node is a root of the cache-sensitive prefix tree.
15. A computing system comprising: a column-oriented database
system; a dictionary-based storage unit specifying a mapping
between a plurality of variable-length string values and a
plurality of integer codes in the column-oriented database system;
shared-leaves data structures that hold data of the dictionary
storage unit in sorted order in their leaves; a processor in
communication with the dictionary-based storage unit, the processor
operable to encode the plurality of variable-length string values
to the plurality of integer codes and decode the plurality of
integer codes to the plurality of variable-length string values
using the shared-leaves data structures.
16. The system of claim 15, wherein the shared-leaves data
structures include an encode index to encode the plurality of
variable-length string values and a decode index to decode the
plurality of integer codes.
17. The system of claim 16, wherein the encode index supports
propagation of the plurality of variable-length string values to
the shared leaves, lookup of the plurality of integer codes, and
generation of a second plurality of integer codes if a subset of
the plurality of integer codes is not found during lookup.
18. The system of claim 16, further comprising a cache-sensitive
array trie index or a cache-sensitive prefix tree index used on top
of the shared-leaves data structures as the encode index.
19. The system of claim 15, wherein the cache-sensitive array trie
index comprises an array to store the plurality of variable-length
string values.
20. The system of claim 15, wherein the cache-sensitive prefix tree
index comprises a node that includes shortest prefixes that enable
propagation of the plurality of variable-length string values to
child nodes of the cache-sensitive prefix tree.
Description
TECHNICAL FIELD
[0001] Embodiments of the invention generally relate to the
software arts, and more specifically, to data structures that
support an order-preserving dictionary compression for string
attributes with a large domain size that may change over time.
BACKGROUND
[0002] In the field of computing, a database management system
(DBMS) is a set of software programs that controls the
organization, storage, management, and retrieval of data in a
database storage unit. Traditionally, the DBMS is a row-oriented
database system; however, there are database systems that are
column-oriented. The column-oriented database systems store their
content by column rather than by row. This may have advantages for
databases, where the aggregates are computed over large numbers of
similar data items. A column-oriented implementation of a DBMS
would store attributes of a given column in sequence, with the
column values for the same column stored in sequence, with the end
of one column followed by the beginning of the next column.
Column-oriented database systems may be more efficient when an
aggregate has to be computed over many rows but only for a smaller
subset of all columns of data. This may be so at least because,
reading that smaller subset of data can be faster than reading all
data. Column-oriented database systems may also be more efficient
when new values of a column are supplied for all rows at once,
because that column data can be written efficiently and can replace
old column data without interfering in any other columns for the
rows.
SUMMARY
[0003] Methods and systems are described that involve data
structures that support order-preserving dictionary compression of
variable-length string attributes where the domain size is large or
not known in advance. In one embodiment, the method includes
propagating a plurality of string values to the compressed data of
a shared-leaves structure of a dictionary via an encode index. A
plurality of order-preserving integer codes is obtained for the
plurality of string values via a lookup operation. If a subset of
the plurality of integer codes was not found during the obtainment,
a subset of the plurality of string values for which the subset of
the plurality of integer codes was not found is inserted into the
shared-leaves structure. The method also includes generating the
subset of the plurality of integer codes for the corresponding
subset of the plurality of string values. Finally, a list of the
order-preserving plurality of integer codes is provided, wherein
the list includes the generated subset of the plurality of integer
codes as well.
[0004] In one embodiment, the system includes a column-oriented
database system and a dictionary-based storage unit specifying a
mapping between a plurality of variable-length string values and a
plurality of integer codes in the column-oriented database system.
Further, the system includes shared-leaves data structures that
hold the data of the dictionary-based storage unit in sort order in
their leaves. In addition, a processor in communication with the
dictionary-based storage unit is included, wherein the processor is
operable to encode the plurality of variable-length string values
to the plurality of integer codes and decode the plurality of
integer codes to the plurality of variable-length string values
using the shared-leaves data structures.
[0005] These and other benefits and features of the embodiments of
the invention will be apparent upon consideration of the following
detailed description of preferred embodiments thereof, presented in
connection with the following drawings in which like reference
numerals are used to identify like elements throughout.
BRIEF DESCRIPTION OF THE FIGURES
[0006] The invention is illustrated by way of example and not by
way of limitation in the figures of the accompanying drawings in
which, like references, indicate similar elements. It should be
noted that references to "an" or "one" embodiment in this
disclosure are not necessarily to the same embodiment, and such
references mean at least one.
[0007] FIG. 1 is a diagram of an embodiment for order-preserving
dictionary compression of variable-length string attributes.
[0008] FIG. 2 is a diagram that shows a direct, an indirect, and a
shared-leaves approach for indexing of the attributes of table
T.
[0009] FIG. 3 is a flow diagram of an embodiment that shows how the
shared leaves approach supports data loading and query processing
inside a dictionary.
[0010] FIG. 4 is a diagram of an embodiment that shows the
all-bulked and hybrid strategies.
[0011] FIG. 5 is an example of a memory layout of a leaf structure
for variable-length string values, according to an embodiment of
the invention.
[0012] FIG. 6 is a diagram of an embodiment for encoding data using
a CS array trie.
[0013] FIG. 7 is a diagram of an example of a CS prefix tree,
according to an embodiment.
[0014] FIG. 8 is a flow diagram of an embodiment of a bulk load
procedure of a CS prefix tree from a given leaf level.
[0015] FIG. 9 is a diagram of results from an experiment that shows
performance and memory overhead of the leaf structure, according to
an embodiment.
[0016] FIG. 10 is a diagram of results from an experiment that
shows lookup and update costs of encoding indexes, according to an
embodiment.
[0017] FIG. 11 is a diagram of results from an experiment that
shows scalability of the dictionary, according to an
embodiment.
[0018] FIG. 12 is a schematic diagram of an example of a computer
system, according to an embodiment of the invention.
DETAILED DESCRIPTION
[0019] Embodiments of the invention relate to data structures that
support an order-preserving dictionary compression for string
attributes with a large domain size that is likely to change over
time. It is believed that the column-oriented database systems
perform better than the traditional row-oriented database systems
on analytical workloads. Lightweight compression schemes for
column-oriented database systems may enable query processing on
compressed data and thus improve query processing performance.
Dictionary encoding replaces variable-length attribute values with
shorter, fixed-length integer codes. To compress column data in
this way, existing column stores usually create a dictionary array
of distinct values and then store each attribute value as an index
into that array. Dictionaries may be used in column stores if the
domain size is small.
[0020] Bit packing can be used on top of a dictionary to compress
the data further by calculating the minimal number of bits needed
to code the maximal index into the dictionary. Bit packing is
useful if the size of the domain is stable or known in advance, but
in application scenarios the domain size may increase over time. If
the domain size is not known in advance, a column store may analyze
the first bulk load of data to find out the current domain size of
a given attribute and then derive the minimal number of bits (for
bit packing). If subsequent bulk loads contain new values, all the
previously loaded data can be decoded and then encoded again with
the new load using more bits.
[0021] Column stores often use order-preserving compression schemes
to speed up expensive query operations because the operations can
then be executed directly on the encoded data. However, such
compression schemes generate either variable-length codes that are
expensive or fixed-length codes that are difficult to extend. For
large dictionaries where the domain size is not known in advance,
sorted arrays and fixed-length integer codes for indexes are too
expensive.
[0022] FIG. 1 is a diagram of an embodiment for order-preserving
dictionary compression of variable-length string attributes. In an
embodiment, data structures are defined that support
order-preserving dictionary compression of variable-length string
attributes where the domain size is large or not known in advance.
The generated integer codes have a fixed length and do not use bit
packing. A dictionary may be defined as a table that specifies a
mapping between string values and integer codes. Index structures
for access to the dictionary may be defined that support: data
loading 110, query compilation 120, and query execution 130.
Process 100 begins with data loading 110. Usually, data is loaded
bulk-wise into a data warehouse. In an embodiment, the dictionary
supports encoding of bulks of string values using integer codes.
The encoding may consist of the following operations: 1) bulk
lookup of the integer codes for the string values in the
dictionary; and 2) bulk insertion of new string values plus
generation of order-preserving integer codes for the new values.
FIG. 1 shows how two bulks of product data, bulk 105 and bulk 115,
are loaded in a string dictionary (e.g., 125) and encoded using
integer codes 135.
[0023] Further, process 100 includes query compilation 120. To
execute analytical queries on encoded data, it is necessary to
rewrite the query predicates. A predicate is a phrase template that
describes a property of objects or a relationship among objects
represented by the variables. For example, string dictionary 125
includes a list of string values: Whole Milk--Gallon, Whole
Milk--Quart, etc, wherein "Whole Milk" is the predicate of these
strings. Query compilation 120 involves rewriting a string constant
in an equality predicate (e.g., p_name="Whole Milk--Gallon") or in
a range predicate (e.g., p_name.gtoreq."Whole Milk--Gallon") with
the corresponding integer code. An order-preserving encoding scheme
allows the string constants of equality and range predicates to be
replaced by integer codes, and prefix predicates (e.g.,
p_name="Whole Milk*") to be mapped to range predicates. For
example, original query 140 is rewritten in query 145, by rewriting
string constant p_name with prefix predicate "Whole Milk*" into a
range predicate 32100.gtoreq.p_name.gtoreq.32000. In an embodiment,
the string dictionary 125 supports lookup operations to rewrite
string constants as well as string prefixes. Process 100 also
includes query execution 130. During query execution 130, encoded
query results 150 are decoded using the dictionary 125. In an
embodiment, the string dictionary 155 supports decoding of the
encoded query results 150 given as a list of integer codes to
generate query results as string values 160. The encoded query
results 150 are decoded to a list of non-encoded query results 160
that represents string values 105.
[0024] In an embodiment, the string dictionary is a table T with
two attributes: T=(value, code). Table T defines a mapping of
variable-length string values (defined by the attribute value) to
fixed-length integer codes (defined by the attribute code) and vice
versa. The dictionary supports the following operations for
encoding and decoding string values and to enable rewrite of the
query predicates: 1) encode: values.fwdarw.codes; 2) decode:
codes.fwdarw.values; 3) lookup: (value, type).fwdarw.code; and 4)
lookup: prefix.fwdarw.(mincode, maxcode). The
"encode:.fwdarw.values codes" operation is used during data loading
110 to encode the data of a string column (i.e., the values) with
the corresponding integer codes (i.e., the codes). This operation
includes the lookup of codes for those strings that are already in
the dictionary and the insertion of new string values as well as
the generation of order-preserving codes for these new values. The
"decode: codes.fwdarw.values" operation is used during query
processing 130 to decode bulk results using the corresponding
string values. The "lookup: (value, type).fwdarw.code" operation is
used during query compilation 120 to rewrite a string constant in
an equality-predicate (e.g., p_name="Whole Milk--Gallon") or in a
range-predicate (e.g., p_name.gtoreq."Whole Milk--Gallon") with the
corresponding integer code. The parameter "type" specifies whether
a dictionary should execute an exact-match lookup or return the
integer code for the next smaller string value. The "lookup:
prefix.fwdarw.(mincode, maxcode)" operation is used during query
compilation 120 to rewrite the prefix of a prefix-predicate (e.g.,
p_name.fwdarw."Whole Milk*") with the corresponding integer ranges
(i.e., the mincode and the maxcode).
[0025] FIG. 2 is a diagram that shows a direct, an indirect, and a
shared-leaves approach for indexing of the attributes of table T.
Table T can support the above operations for encoding and decoding
string values and to enable rewrite of the query predicates by
building indexes for both attributes of the table T (value and
code). In an embodiment, the indexing of the attributes value and
code can be direct 210 or indirect 220. Structure 200 shows the
direct 210, indirect 220, and shared-leaves 230 approach. In the
direct approach 210, two indexes for encoding and decoding are
created that hold the table data directly in their leaves (e.g.,
encode index 205 and decode index 215). The table does not need to
be kept in the main memory since the data is stored in the indexes,
but the table data is held redundantly in the two indexes, which
wastes memory space. In the indirect approach 220, the encode index
and decode index for encoding and decoding, such as encode index
205 and decode index 215, are created to hold only references to
the table data 225 (e.g., using a row identifier "rid" 235). The
table may be kept in main memory instead being written to an
external drive. Although the indexes are smaller, the extra level
of indirectness may result in higher cache miss-rates. In another
embodiment, the two indexes for encoding and decoding share the
same leaves (e.g., shared-leaves 230 approach), which hold the
table data directly in their leaves but avoid the redundancy of the
direct indexing 210 approach. Thus, the shared leaves also avoid
the additional level of indirectness of the indirect indexing
approach 220.
[0026] Since the string dictionary uses an order-preserving
encoding scheme, the string values and the integer codes in table T
follow the same sorting order. As both attribute values of table T
can be kept in sorting order inside the leaves, the leaves can
provide access paths for both lookup directions (i.e., for the
encoding and decoding) using a standard search method for sorted
data (e.g., binary search or interpolation search). Moreover, as is
the case for direct indexes, using the shared-leaves for indexing
the dictionary means that table T does not have to be kept
explicitly in main memory because the leaves hold all the data of
table T.
[0027] FIG. 3 is a flow diagram of an embodiment that shows the
shared leaves approach for data loading and query processing inside
a dictionary. An encode index may be used to encode a list of
string values. At block 305 of process 300, the encode index
propagates (e.g., broadcasts, passes along, etc.) the string values
to the corresponding leaves (e.g., as shown in FIG. 6 or FIG. 7).
At block 310, lookup of the integer code for each single string
value is performed. In one embodiment, the lookup operation may be
performed via a standard search algorithm inside a leaf. The
decoding of a list of integer codes works similarly wherein the
decode index propagates the integer codes down to the corresponding
leaves. At decision block 315, the process checks if for each
string value there is an integer code. If the corresponding integer
codes for all string values are found, then the process continues
at block 330. At block 330, a list of integer codes is returned by
the lookup operation (i.e., a result list). If some integer codes
for string values are not found by the lookup operation on the
encode index, the process continues at block 320. At block 320, the
string values for which there are no integer codes found are
inserted into the dictionary by writing new rows in table T. New
integer codes are generated for these string values, at block 325.
The code range for insertion of new string values is partitioned
into equidistant intervals. The limits of the intervals represent
the new codes. If the range is smaller than the number of new
string values to be inserted, some string values are recoded and
the data using them is updated. The newly generated integer codes
are added to the list of codes returned by the lookup operation at
block 315. The indexes are updated if necessary.
[0028] In an embodiment, the shared leaves also support rewriting
predicates inside a dictionary. For rewriting equality and range
predicates, the encode index propagates the string values to the
corresponding leaves and a search operation on the leaves returns
the integer codes. For rewriting prefix predicates, the encode
index propagates the prefix to the leaves containing the minimal
and maximal string values for the prefix; the leaves map the
minimal and maximal strings to the integer codes.
[0029] The data structures of the dictionary (i.e., leaves and
indexes) are optimized for encoding or decoding bulks and are
cache-aware. The operations, encoding and decoding, are easy to
parallelize. The leaf structure differs from the index structure.
All structures, leaf structures and index structures, reside in
memory. The leaf structure holds the string values and the integer
codes in sorting order. A leaf supports the encoding of
variable-length string values and supports efficient bulk loads and
bulk updates. The indexes for encoding and decoding keep the keys
in sort order for efficient lookup over the sorted leaves. The
encode index provides propagation of string constants and string
prefixes to the leaves.
[0030] In addition to the lookup of the integer codes for string
values that are already a part of the dictionary, it might be
necessary to insert new string values into the dictionary (e.g.,
update the leaves as well as the both indexes for encoding and
decoding) and generate new order-preserving codes for these values.
In an embodiment, the lookup and insert operations are combined
into one operation. The following strategies can support this
approach: all-bulked that updates the encode and decode indexes
after generation of any new integer codes and hybrid approach that
updates the encode index during propagation of the string
values.
[0031] FIG. 4 is a diagram of an embodiment that shows the
all-bulked and hybrid strategies. Process 400 shows the main steps
of the all-bulked 401 strategy and the hybrid 402 strategy.
All-bulked 401 strategy begins at block 410. At block 410, the
string values are propagated to the corresponding leaves using an
encode index. The integer codes for the string values in the leaves
are looked up. At block 420, any new string values are inserted
into the leaves. The leaf level is reorganized in response to the
inserted new values (e.g., a leaf level is created where all leaves
are filled up to the maximal leaf size). At block 430, integer
codes for the new values are generated. At block 440, a new encode
index and a new decode index are bulk loaded from the updated leaf
level. Hybrid 402 strategy includes the following steps. At block
450, the string values are propagated to the corresponding leaves
using an encode index. The encode index is updated directly, during
propagation of the string values. At block 450, the integer codes
for the string values in the leaves are looked up. At block 470,
any new string values are inserted into the leaves. Integer codes
for the new values are generated. At block 480, a new decode index
is bulk loaded from the updated leaf level. To ensure data
consistency, the indexes and the leaves are locked during data
loading. For query processing, read-only concurrency may be
allowed. During data loading, updated leaves may be written
sequentially to a storage unit.
[0032] FIG. 5 is an example of a memory layout of a leaf structure
for variable-length string values, according to an embodiment of
the invention. Leaf structure 500 may be used in the shared-leaves
approach 230 for encoding and decoding variable-length string
values on a particular platform. The leaf structure 500 keeps
string values as well as their corresponding fixed-length integer
codes sorted and compressed together in one chunk of memory to
increase the cache locality during data loading and lookup
operations. For a lookup using this leaf structure 500, an offset
vector 505 is stored at the end of the leaf that holds references
(i.e., offsets) and the integer codes for all uncompressed strings
of a leaf also in a sorted way. For example, offset 128 at 510 and
code 32000 at 520 are stored in the offset vector 505 for value 16
at 530. The integer codes of the compressed string values are
stored together with the compressed string values in the data
section and not in the offset vector 505 (e.g., code 32050 for
value 17).
[0033] During leaf structure data loading, the string values are
first compressed and then written into the leaf together with their
codes in a forward way (e.g., starting from memory position 0 and
incrementing the position for each new entry). To enable searching
for string values inside a leaf includes, each n-th string (e.g.,
each third string) is stored in an uncompressed way and the
positions of the uncompressed strings are saved as anchors at the
end of a leaf (to be found during search). However, when loading
data into a leaf, the exact size of the compressed string values
may be unknown before all data is written into the leaf. Thus, the
offset vector 505 may be stored from the last memory position in a
reverse way by decrementing the position for each new anchor.
[0034] For bulk lookup, the leaf structure supports one lookup
operation to look up the integer code for a given string value and
another to look up the string value for a given code. To look up
the code for a given string value, an algorithm may be performed
for sequential search over the incrementally encoded values that
does not require decompression of the leaf data. In an embodiment,
the algorithm may be as described with reference to table 1
below.
TABLE-US-00001 TABLE 1 Sequential search of string v on compressed
leaf Algorithm 1: Sequential search of string v on compressed leaf
function SEQUENTIALSEARCH(leaf, start, end, v) v' leaf[start] ---
read string v' at offset start start start + size(v') --- increment
offest by string size prefix_len prefix_len(v, v') --- calculate
common prefix len while start .sub.-- end and prefix_len < |v|
do curr_prefix_len leaf[start] --- get curr. prefix len start start
+ p --- increment offest by prefix-size p = 1 v' leaf[start] start
start + size(v') if curr_prefix_len <> prefix_len then
continue --- prefix of curr. value v' too short/long else if
compare(v', v) > 0 then return -1 --- curr. value v' comes after
search value v end if prefix_len prefix_len + prefix_len(v, v')
start start + c --- increment offest by code-size c = 4 end while
if prefix_len = |v| then return leaf[start - c] --- string v found:
return code else return -1 --- string v not found end if end
function
[0035] The values are incrementally decompressed during the
sequential search when looking up a value for a given code. For
bulk lookup, the lookup probe is sorted to reduce the search
overhead. For the initial bulk load of a leaf with a list of string
values, the string values are sorted first. Then, the leaf data is
written sequentially from the beginning of the leaf and the offset
vector is written in reverse order from the end. If the string
values do not occupy all the memory allocated for them, the offset
vector is moved forward and the unused memory released. For bulk
update, a list of new string values is first sorted and then
inserted into an existing leaf. Then, a sort merge of the new
string values and the existing leaf is performed to create a new
leaf.
[0036] In an embodiment, a cache-conscious index structure may be
used on top of the leaf structure for encoding and decoding. These
indexes support the all-bulked 401 and hybrid 402 strategies. For
the encoding index, a cache-sensitive (CS) version of the Patricia
trie, the CS array trie, that supports the hybrid strategy 402 is
defined. The Patricia trie, or radix tree, is a specialized set
data structure based on the trie (a prefix tree that is an ordered
tree data structure used to store associative array where the keys
are usually strings) that is used to store a set of stings. In
contrast with a regular trie, the edges of a Patricia trie are
labeled with sequences of characters rather than with single
characters. These can be strings of characters, bit strings such as
integers or IP addresses, or generally arbitrary sequences of
objects in lexicographical order. In addition, a new
cache-sensitive version of the prefix B-tree (a tree data structure
that keeps data sorted and is optimized for systems that read and
write large bulks of data), the CS prefix tree, to support the
all-bulked 401 update strategy is defined. As decoding index, a CS
search (CSS) tree may be used. The CSS tree may be created over the
leaves of the dictionary using the minimal integer codes of each
leaf as keys of the index. The CSS tree can be bulk loaded
efficiently bottom-up from the leaves of the dictionary. A CS array
trie may be used as an encode index to propagate string lookup
probes and updates to the leaves. The CS array trie uses
read-optimized cache-aware data structures for the index nodes and
does not decompose the strings completely.
[0037] FIG. 6 is a diagram of an embodiment for encoding data using
a CS array trie. A CS array trie 610 node uses an array instead of
a linked list to store the characters of the indexed string values.
When sequentially inserting single values into a trie, an array is
less efficient than a linked list. But for each bulk insert of new
values into an array trie, the array of a node is expanded only
once. For lookup of a string value of a CS array trie 610, a search
over the characters of a node may be performed, where the array
supports binary search and all characters are stored clustered in
memory. The CS array trie 610 stores a set of strings that have the
same prefix together using the leaf structure. For example, leaf
620 includes a set of strings that have the same "aa" prefix. The
CS array trie leaf 610 stores the complete strings to enable
efficient decoding of integer codes using the same leaves. The
strings are compressed using incremental encoding.
[0038] In an embodiment, the CS array trie may be used to implement
the hybrid update strategy 402 for bulk encoding of string values
during data loading 110, as shown at 630. The string values 640 are
propagated (in preorder) to the leaves using variable buffers
(e.g., buffers 645, 650, and 650) at each trie node to increase
cache locality for lookup. Using buffers at each node, the array of
characters stored at a node grow only once per bulk. This reduces
cache misses. To estimate the expected leaf size, the uncompressed
size of all new strings can be added in a buffer page as well as
the size of their new codes (without eliminating duplicates) to the
current leaf size. When all string values 640 are propagated to
their leaves, new integer codes 660 are generated for the new
string values by analyzing the number of strings inserted between
existing string values 640.
[0039] The CS array trie supports efficient predicate rewrite. For
equality and range predicates, the constants are propagated through
the trie without buffering. For prefix predicates, the prefix is
used to find the minimal and maximal string values that match it.
Propagation of string values from the root of the trie to the
leaves is parallelized. New integer codes are generated in
parallel, without locking any data structures, by determining which
leaves hold contiguous new string values. Lookup of the new string
values can also be parallelized without locking any data
structures.
[0040] FIG. 7 is a diagram of an example of a CS prefix tree,
according to an embodiment. In an embodiment, a CS prefix tree can
be used as an encode index to propagate string lookup probes and
updates. A CS prefix tree node contains the shortest prefixes that
enable the propagation of string values to the child nodes. Instead
of storing a pointer to each child, the CS prefix tree allocates a
contiguous block of memory for nodes and uses offsets to navigate
through the block. This reduces memory consumption and prevents
performance problems due to pointer chasing. To reduce the memory
footprint of the tree further, only the offset to the first child
node 730 is stored explicitly. With nodes of fixed size, offset
arithmetic may be used to calculate the offset to a child node. To
enable fast search over the variable-length keys of a node, the
offsets to the keys are stored in an offset vector at the beginning
of each node. The node size (s) is fixed to allow use of offset
arithmetic for computing the index (i) to the child nodes. For
example, the i-th child of a node can be found at offset o=offset
(first_child)+(i*s). The number of children of a node is variable
because variable-length keys are stored at a node.
[0041] The CS prefix tree can only be bulk loaded bottom-up, so it
is mainly suitable for the all-bulked update strategy 401. To
encode the first bulk load of string values, the string values are
used to build the complete leaf level. A CS array trie may be used
to partition the string values into buckets sorted using multi-key
quick-sort. Then, the sorted string values are used to create and
fill in leaves 720 to the maximum leaf size. From these leaves 720,
a new encode index is bulk loaded bottom-up.
[0042] FIG. 8 is a flow diagram of an embodiment of a bulk load
procedure of a CS prefix tree from a given leaf level. At block
810, starting with the first two leaves, a first shortest prefix is
calculated. The first shortest prefix to distinguish the largest
value of the first leaf and the smallest value of the second leaf.
At block 815, the first shortest prefix is stored in a node of the
tree. Because an offset into the leaf level can be calculated, a
pointer to each child is not stored. At block 820, an offset vector
is written from left to right and the keys are stored from right to
left (since the size of the offset vector is unknown). At block
825, a second shortest prefix is calculated. The second shortest
prefix to distinguish the largest value of the second leaf and the
smallest value of the third leaf, and so on until all leaves are
processed. At block 835, a new node of the tree is started if the
current tree node is full. The index is stored to the first leaf
that will be a child of this new node as an anchor. At block 840,
the nodes are stored contiguously in a memory storage unit. At
block 845, as long as more than one node is created for a given
level of the tree, another level is added on top of existing tree
levels. The new level includes a node storing prefixes
distinguishing the child nodes. This node is the root of the tree.
The offset is stored to this node in the next memory block. Because
the tree is built bottom-up, the nodes are stored in that sequence
in memory.
[0043] For subsequent bulk loads, the existing CS prefix tree may
be used to propagate the string values to the leaves. The string
values are buffered at leaf level and then the existing leaf is
sort-merged with the new string values stored in the buffers. If
the new string values in the buffers and the values of the existing
leaf do not fit into one leaf, another leaf may be created. Query
predicates can be rewritten using the CS prefix tree. For an
equality predicate or a range predicate, a simple lookup can be
performed with the string constants. For a prefix predicate, the
prefix can be used to find the minimal string value that matches
the prefix. The lookup for the prefix finds a leaf containing the
value even if the value is not in the dictionary. From that leaf
on, a sequential search can be executed for the maximum string
value matching the prefix.
[0044] Memory may be allocated in advance in contiguous blocks. The
maximum amount of memory that all the tree nodes need can be
calculated by setting an arbitrary limit on the maximum length of
the keys in the tree. Then, the minimum number of keys that fit in
one node is calculated and hence the maximal number of nodes needed
to store the data. Using a mix of pointers and offset arithmetic
identifies the correct child and thus allows use of multiple blocks
of memory. A CS prefix tree may be more expensive to build than a
CS array trie because the data is first sorted and then loaded
bottom-up. But the CS prefix tree performs better than the CS array
trie for lookup workloads.
[0045] FIG. 9 is a diagram of results from an experiment that shows
performance and memory overhead of the leaf structure, according to
an embodiment. The experiment includes two workloads so that each
fit into a leaf of size 16 MB. The workloads use different fixed
string lengths and thus represent different numbers of strings. The
workloads are distinct unsorted (without skew) to represent the
worst case for lookup operations. A workload is loaded onto a set
of leaves holding the data in sort order. The used leaf sizes vary
from 64 KB to 16 MB. The costs of the different workloads with
different leaf sizes are measured without the overhead of an
encoding and decoding index. Also, the execution time and cache
misses when bulk loading the leaves 910 and executing lookups for
encoding 920 and decoding 930 are measured. To load the leaves,
first the workloads are sorted and each leaf is bulk loaded up to
its maximal size, then the integer codes are generated for the
leaves. Size of 8 bytes is used for the integer code to find the
expected memory consumption for encoding attributes with a large
domain size. To measure the lookup performance of the leaf
structures, each string value of the workloads is assigned to a
leaf using a buffer and then the code for each string is looked up
in the individual buffers. Finally, the encoded workload is used to
execute the lookup for decoding the integer codes again.
[0046] In the experiment, a 16 MB leaf structure is compared to two
cache-sensitive read-optimized index structures using two different
workloads. For encoding the string values, the leaf structure is
compared to a compact-chain hash table (i.e., bulk lookup 920). For
decoding integer codes, the leaf structure is compared to a CSS
tree (i.e., bulk lookup 930). The result shows that the optimal
leaf structure size is about 512 KB (medium) and the performance of
the leaf structure is comparable to read-optimized index structures
yet uses less memory.
[0047] FIG. 10 is a diagram of results from an experiment that
shows lookup and update costs of encoding indexes, according to an
embodiment. The costs for encoding workloads that produce different
update patterns on the indexes are measured. First, a dictionary
with 10 million strings is bulk loaded (Bulk 1) and then another
bulk load (Bulk 2) of 10 million strings representing a certain
update pattern is encoded. All workloads consist of string values
with a fixed length of 20 characters. The maximum leaf size is set
at 512 KB. In the experiment, five different update patterns are
used. 1) No-updates: Bulk 2 contains no new strings; 2) Interleaved
10: Bulk 2 contains 10% new string values where each tenth string
in sort order is new; 3) Interleaved 50: Bulk 2 contains 50% new
string values where every other string in sort order is new; 4)
Interleaved 100: Bulk 2 contains 100% new string values, each
inserted between two values of Bulk 1; and 5) Append: Bulk 2
contains 100% new string values, all inserted after the last string
of Bulk 1. The experiment shows that the times for loading Bulk 1
and then Bulk 2 for the different update patterns were similar for
CS array tries and CS prefix trees but much longer for list
trie.
[0048] FIG. 11 is a diagram of results from an experiment that
shows scalability of the dictionary, according to an embodiment.
The experiment tests the performance of the data structures for
unsorted workloads of different sizes. The results show that the
times required to encode the workloads using different encoding
indexes for the dictionary showed linear scaling with the number of
strings in the workload. After encoding, a decode index is bulk
loaded from the leaves of the encoding indexes and the encoded
workloads are decoded. Again, the times showed linear scaling. To
check the effects of parallelization on the CS array trie, the
number of threads was varied and the results show that performance
with 8 threads was better than with 16 threads due to the overhead
for thread synchronization.
[0049] Lightweight compression schemes can improve the query
processing performance of column-oriented database systems. In one
such scheme, a dictionary replaces long (variable-length) values
with shorter (fixed-length) integer codes. To improve performance
further, column stores can use order-preserving compression
schemes. New data structures may be used to support
order-preserving dictionary compression for variable-length string
attributes with a large domain size that can change over time. A
dictionary can be modeled as a table mapping string values to
arbitrary integer codes. A new indexing approach may be used for
efficient access to such a dictionary using compressed index data.
The data structures are at least as fast as other data structures
for dictionaries but occupy less memory.
[0050] FIG. 12 is a schematic diagram of an example of a computer
system, according to an embodiment of the invention. System 1200
can be used for the operations described in association with the
FIG. 3 according to one implementation. System 1200 includes a
processor 1210, a main memory 1220, a storage unit 1230, and an
input/output device 1240. Each of the components 1210, 1220, 1230,
and 1240 are interconnected using a system bus 1250.
[0051] The processor 1210 is capable of processing instructions for
execution within the system 1200. The processor is in communication
with the main memory store 1220. Further, the processor is operable
to execute operations 1280 stored in the main memory 1220, such as
data loading 110, query compilation 120, and query execution 130.
In one embodiment, the processor 1210 is a single-threaded
processor. In another embodiment, the processor 1210 is a
multi-threaded processor. The processor 1210 is capable of
processing instructions stored either in main memory 1220 or on the
storage device 1230, to display graphical information for a user
interface on the input/output device 1240.
[0052] The main memory 1220 stores information within the system
1200. In one implementation, the main memory 1220 is a
machine-readable medium. In an embodiment, the main memory 1220
stores order-preserved compressed data in a column-oriented format.
Main memory 1220 stores a dictionary 1260. Dictionary 1260 is used
for encoding and decoding of the compressed data, as represented by
index 1270. The encode index and decode index contain shared-leaves
data structures that hold data in sorted order in their leaves.
[0053] The storage device 1230 is capable of providing mass storage
for the system 1200. In one implementation, the storage device 1230
is a computer-readable medium. In alternative embodiments, the
storage device 1230 may be a floppy disk device, a hard disk
device, an optical disk device, or a tape device.
[0054] The input/output device 1240 is used to trigger or initiate
input/output operations 1280 for the system 1200. In one
implementation, the input/output device 1240 includes a keyboard
and/or pointing device. In another implementation, input/output
device 1240 includes a display unit for displaying graphical user
interfaces.
[0055] Elements of embodiments may also be provided as a tangible
machine-readable medium (e.g., computer-readable medium) for
tangibly storing the machine-executable instructions. The tangible
machine-readable medium may include, but is not limited to, flash
memory, optical disks, CD-ROMs, DVD ROMs, RAMs, EPROMs, EEPROMs,
magnetic or optical cards, or other type of machine-readable media
suitable for storing electronic instructions. For example,
embodiments of the invention may be downloaded as a computer
program, which may be transferred from a remote computer (e.g., a
server) to a requesting computer (e.g., a client) via a
communication link (e.g., a modem or network connection).
[0056] It should be appreciated that reference throughout this
specification to "one embodiment" or "an embodiment" means that a
particular feature, structure or characteristic described in
connection with the embodiment is included in at least one
embodiment of the present invention. Therefore, it is emphasized
and should be appreciated that two or more references to "an
embodiment" or "one embodiment" or "an alternative embodiment" in
various portions of this specification are not necessarily all
referring to the same embodiment. Furthermore, the particular
features, structures or characteristics may be combined as suitable
in one or more embodiments of the invention.
[0057] In the foregoing specification, the invention has been
described with reference to the specific embodiments thereof. It
will, however, be evident that various modifications and changes
can be made thereto without departing from the broader spirit and
scope of the invention as set forth in the appended claims. The
specification and drawings are, accordingly, to be regarded in an
illustrative rather than a restrictive sense.
* * * * *