Efficient XML Storage based on DTMEfficient XML Storage based on DTM

for Readfor Read--oriented Workloadsoriented Workloads

Graduate School of Information Science,

Nara Institute of Science and Technology

Makoto Yui

Jun Miyazaki, Shunsuke Uemura, Hirokazu Kato

International Workshop on Advanced Storage Systems 2007 (ADSS 2007)

2

Outline

� Motivation� Related work� Document Table Model (DTM)� XML storage based on DTM

– System Overview– Physical Layout– Buffer Management

� Experimental Evaluation� Conclusions

3

Motivation - Backgrounds

� Past research topics in XML data management

Towards efficient XML data processing

Labeling and indexing XML trees

Join processing

Indexing on paths

� Dewey ordering

� Ordpaths

� XR-Trees

� XRel

� Index Fabric

� Structural-joins

� Twig-joins

Internaldata model Buffer management

Physical data layout� Natix

� Relational

� Hybrid (SystemRX)

� Tree (e.g., DOM)

Well studied

topics

Less studied

topics

Our focus

4

Motivation – Design goals

� Design an XML storage scheme optimizedfor read-oriented workload

– Node-level update is not always required

– Updating capabilities have more or less drawbacks

� Focus on iterative XQuery processingin which an operator tree consists of iterators

The design of read-oriented database has been suggested for relational databases but not studied for XML databases yet

– Ideal XML storage scheme depends on the processing model

(e.g., tuple-at-a-time or set-at-a-time)

How the data access is achieved for each of the two processing model ?

Norman May, et al., Index vs. Navigation in XPath Evaluation. XSym 2006

The reason we selected this model is

set-at-a-time processing of XQuery

involves lots of joins, and it causes performance deteriation.

5

site

regions regions

site

auction

regions

site/regions

Tree traversal of set-at-a-time processing

Context node

Traversed in a breadth-first search manner

sites under actions

are selected

regions under sites

are selectedEach edges are connected by join

on nodes’ identifiers

6

site

regions regions

site

auction

regions

site/regions

Tree traversal of tuple-at-a-time processing

Context node

Traversed in a depth-first search manner

It is as same manner as a document ordering

• Each two nodes are connected by links

• Path processing is achieved by

navigating link edges

7

Motivation

� Design an XML storage scheme optimizedfor read-oriented workload

� Focus on iterative XQuery processingin which an operator tree consists of iterators

We examined actual data access patternswhen evaluating XQuery queries in order to design the suitable data layout

8

Outline

� Motivation� Related Work� Document Table Model (DTM)� XML Storage based on DTM (pDTM)

– System Overview– Pyshical Layout– Buffer Management

� Experimental Evaluation� Conclusions

9

Natix (University of Mannheim, Germany)T. Fiebig, et.al. Anatomy of a Native XML Base Management System.

VLDB Journal, 2002.

f1

f4

f8f7f3f2 f5 f6

h1 h2

p3p2p1

r1

r2 r3 r4

P Proxy Node h Helper aggregate nodes:

Related work (1) storing scheme based on subtrees

� Pros� Cons

Effective for breadth-first traversals

Not effective for depth-first traversals

Allocates a page based on subtrees

10

Related work (2) storing scheme based on a schema

OrientStore (Renmin University, China)X. Meng et.al. OrientStore: A Schema Based Native XML Storage System.VLDB 2003.

� Pros� Cons

• Effective for path processing

• Schema information is required

• Not effective for serialization and

string-value calculationClusters records according to

a semantic segmentation

11

Outline

� Motivation� Related work� Document Table Model (DTM)� XML Storage based on DTM (pDTM)

– System Overview– Pyshical Layout– Buffer Management

� Experimental Evaluation� Conclusions

12

Document Table Model (DTM)

� Originally used by Apache Xalan XSLT processor

� Expresses an XML document as a table form

DTM table consists of primitive data types

CID

--9107--56-NEXTSIB

1771123210PARENT

ETEETETEEEEvent

10987654321

CID

--9107--56-NEXTSIB

1771123210PARENT

ETEETETEEEEvent

10987654321E

E T E

E E

T

E T

E

1

2

3

4

5

6 7

8 9

10

0

DOM has object footprints(e.g., object instantiation and memory consumptions)

DTM can avoid such object footprints

An XML tree structure can be represented as a table

by using link values

13

AsiaNorthAmerica

Africa…

縈緹緱CID

綋綋繮緱緫繇綋綋縕縠綋NEXTSIB

緱繇繇緱緱緹縈緹緱緫PARENT

翎聱翎翎聱翎聱翎翎翎TYPE

Pyshical

Model

Abstract

Model

Page in

record page

#1 site

#2 region

#3 ..

DTMs

StringChunk

PageBuffer QNameTable

Cache miss

System Overview

14

Pyshical layout

� Before designing physical layout of XML documents, we analyzed actual data access patterns of XQuery queries.

Analyzing data access patterns

In general,

Pages are required in the document-orderSequential accesses are frequently appeared

15

Access pattern analysis of XMark queries (1)

Required Page

Access count

This plot explains that

28,000th data access

requires a page #22,000.

page #0 is assigned to

nodes nearby the root

page #40,000 is assigned to

nodes nearby end of the document

Document order is reflected

to the Y-axis

pages are required in a document order

from the lower left to the upper right.

16

Access pattern analysis of XMark queries (2)

Recall that we claimed that tuple-at-a-time

processing of XPath queries, in general, traverses

XML-tree according to the document-ordering.

Note that the overall tendency is not restricted to XMark queries

but also other benchmark queries.

XQuery FLWR queries also shows

the similar access patterns with

few exceptions where queries contains

lots of backward axes.

17

Pyshical layout

� Pages are required in a document-order� Sequential accesses are frequently appeared

Prefetching is effective

The prefetching entries can compete

for hot cache entries

We conducted informed prefetching with

scan-resistant buffer management

Scan-resistant buffer management

Is also effective to sequential scans in XML query processing

Document-ordered block allocation is suitable

18

Outline

� Motivation� Related work� Document Table Model (DTM)� XML Storage based on DTM (pDTM)

– System Overview– Pyshical Layout– Buffer Management

� Experimental Evaluation� Conclusions

19

Experimental evaluation

� Compared to Natix version 2.1.1 where XMark SF = 5 and SF = 10

� Experimental settings

SF = 5 SF = 10

Elapsed time

in a log scale

Query number

Today’s normal PC setting

20

XMark SF=10

Queries whose outputs are large

Has following-sibling Contains “//”

Our method is effective for

IO-intensive workloads

Natix showed better performance

for breadth-first traversals

(e.g., following-sibling)

21

Conclusions

� Proposed an efficient XML storage schemebase on DTM for iterative XQuery processing

� Our approach is effective for IO-intensive workloadssuch as queries including ‘//’.

� Document-ordered block allocation� Informed prefetching and scan-resistant caching

Summary

Future work� Automatic database tuning based on online analysis of data access patterns (e.g., buffer replacement policy and prefetching)

22

Thank you for your attention!

Questions?

23

Problem in XML-Relational mapping

<student><id>0551133</id><name>ichiro</name>

</student>

XML

Student

id name

XML Tree

A

B C

Hard disk

Relational Table

contentparentid

--A

0551133AB

ichiroA郅

Mapping

24

綋

綋

綋

R

緫縈緹緱羑翮羿

綋綋繮綋繇綋綋縑縠綋耔翎肫聱聠翮罡

緱繇繇緱緱緹縈緹緱緫耦罓耼翎耔聱

翎聱翎翎聱翎聱翎翎翎聱胗耦翎

縈

縑

緹

翎

Updating facilities and versioning

When accessing to a record,

Logical address Physical address

For updating facilities, we change this method as follows:

Logical address Physical address

Address conversion table

25

Buffer management (XMark Q10 as an example)

緫繢緫綆縠繇縈緱繢縕綌縕縠緹耺

縕縠繇綆繇緫緹緱綆繮緱繮綆縕繢縠緹緱緱綌繢縈耎耼聵

膴莆舡舡舃荃糫荃舃苆芎膄臏舃芮舃芷荕

荕芾荕膄芎糫荃舃膄臹糫膴芎芾臏艏荑

翎芎膄苆荑舃臹糫荕艆芮舃糫絓芮荑舃臏絙

14.2% speedups

20% of original document is returned

as the result

26

27

Rows

Blocks

Read in Pages

accessed row

Pages

dynamically

configured

nullRow Skeleton

Pyshical layout

28

DC/SD Normal TC/SD Normal

Access pattern analysis of XBench queries

29

#include <sys/mman.h>

void *mmap(void *start, size_t length, int prot, int flags,int fd, off_t offset);

int munmap(void *start, sizt_t length);

Memory Mapped DTM

We present a memory mapped scheme

extending the DTM model, it has boost

the performance significantly.

Efficient XML Storage based on DTMEfficient XML Storage based on DTM

for Readfor Read--oriented Workloadsoriented Workloads

Graduate School of Information Science,

Nara Institute of Science and Technology

Makoto Yui

Jun Miyazaki, Shunsuke Uemura, Hirokazu Kato

International Workshop on Advanced Storage Systems 2007 (ADSS 2007)

2

Outline

� Motivation� Related work� Document Table Model (DTM)� XML storage based on DTM

– System Overview– Physical Layout– Buffer Management

� Experimental Evaluation� Conclusions

3

Motivation - Backgrounds

� Past research topics in XML data management

Towards efficient XML data processing

Labeling and indexing XML trees

Join processing

Indexing on paths

� Dewey ordering

� Ordpaths

� XR-Trees

� XRel

� Index Fabric

� Structural-joins

� Twig-joins

Internaldata model Buffer management

Physical data layout� Natix

� Relational

� Hybrid (SystemRX)

� Tree (e.g., DOM)

Well studied

topics

Less studied

topics

Our focus

4

Motivation – Design goals

� Design an XML storage scheme optimizedfor read-oriented workload

– Node-level update is not always required

– Updating capabilities have more or less drawbacks

� Focus on iterative XQuery processingin which an operator tree consists of iterators

The design of read-oriented database has been suggested for relational databases but not studied for XML databases yet

– Ideal XML storage scheme depends on the processing model

(e.g., tuple-at-a-time or set-at-a-time)

How the data access is achieved for each of the two processing model ?

Norman May, et al., Index vs. Navigation in XPath Evaluation. XSym 2006

The reason we selected this model is

set-at-a-time processing of XQuery

involves lots of joins, and it causes performance deteriation.

5

site

regions regions

site

auction

regions

site/regions

Tree traversal of set-at-a-time processing

Context node

Traversed in a breadth-first search manner

sites under actions

are selected

regions under sites

are selectedEach edges are connected by join

on nodes’ identifiers

6

site

regions regions

site

auction

regions

site/regions

Tree traversal of tuple-at-a-time processing

Context node

Traversed in a depth-first search manner

It is as same manner as a document ordering

• Each two nodes are connected by links

• Path processing is achieved by

navigating link edges

7

Motivation

� Design an XML storage scheme optimizedfor read-oriented workload

� Focus on iterative XQuery processingin which an operator tree consists of iterators

We examined actual data access patternswhen evaluating XQuery queries in order to design the suitable data layout

8

Outline

� Motivation� Related Work� Document Table Model (DTM)� XML Storage based on DTM (pDTM)

– System Overview– Pyshical Layout– Buffer Management

� Experimental Evaluation� Conclusions

9

Natix (University of Mannheim, Germany)T. Fiebig, et.al. Anatomy of a Native XML Base Management System.

VLDB Journal, 2002.

f1

f4

f8f7f3f2 f5 f6

h1 h2

p3p2p1

r1

r2 r3 r4

P Proxy Node h Helper aggregate nodes:

Related work (1) storing scheme based on subtrees

� Pros� Cons

Effective for breadth-first traversals

Not effective for depth-first traversals

Allocates a page based on subtrees

10

Related work (2) storing scheme based on a schema

OrientStore (Renmin University, China)X. Meng et.al. OrientStore: A Schema Based Native XML Storage System.VLDB 2003.

� Pros� Cons

• Effective for path processing

• Schema information is required

• Not effective for serialization and

string-value calculationClusters records according to

a semantic segmentation

11

Outline

� Motivation� Related work� Document Table Model (DTM)� XML Storage based on DTM (pDTM)

– System Overview– Pyshical Layout– Buffer Management

� Experimental Evaluation� Conclusions

12

Document Table Model (DTM)

� Originally used by Apache Xalan XSLT processor

� Expresses an XML document as a table form

DTM table consists of primitive data types

CID

--9107--56-NEXTSIB

1771123210PARENT

ETEETETEEEEvent

10987654321

CID

--9107--56-NEXTSIB

1771123210PARENT

ETEETETEEEEvent

10987654321E

E T E

E E

T

E T

E

1

2

3

4

5

6 7

8 9

10

0

DOM has object footprints(e.g., object instantiation and memory consumptions)

DTM can avoid such object footprints

An XML tree structure can be represented as a table

by using link values

13

AsiaNorthAmerica

Africa…

321CID

--9107--56-NEXTSIB

1771123210PARENT

ETEETETEEETYPE

Pyshical

Model

Abstract

Model

Page in

record page

#1 site

#2 region

#3 ..

DTMs

StringChunk

PageBuffer QNameTable

Cache miss

System Overview

14

Pyshical layout

� Before designing physical layout of XML documents, we analyzed actual data access patterns of XQuery queries.

Analyzing data access patterns

In general,

Pages are required in the document-orderSequential accesses are frequently appeared

15

Access pattern analysis of XMark queries (1)

Required Page

Access count

This plot explains that

28,000th data access

requires a page #22,000.

page #0 is assigned to

nodes nearby the root

page #40,000 is assigned to

nodes nearby end of the document

Document order is reflected

to the Y-axis

pages are required in a document order

from the lower left to the upper right.

16

Access pattern analysis of XMark queries (2)

Recall that we claimed that tuple-at-a-time

processing of XPath queries, in general, traverses

XML-tree according to the document-ordering.

Note that the overall tendency is not restricted to XMark queries

but also other benchmark queries.

XQuery FLWR queries also shows

the similar access patterns with

few exceptions where queries contains

lots of backward axes.

17

Pyshical layout

� Pages are required in a document-order� Sequential accesses are frequently appeared

Prefetching is effective

The prefetching entries can compete

for hot cache entries

We conducted informed prefetching with

scan-resistant buffer management

Scan-resistant buffer management

Is also effective to sequential scans in XML query processing

Document-ordered block allocation is suitable

18

Outline

� Motivation� Related work� Document Table Model (DTM)� XML Storage based on DTM (pDTM)

– System Overview– Pyshical Layout– Buffer Management

� Experimental Evaluation� Conclusions

19

Experimental evaluation

� Compared to Natix version 2.1.1 where XMark SF = 5 and SF = 10

� Experimental settings

SF = 5 SF = 10

Elapsed time

in a log scale

Query number

Today’s normal PC setting

20

XMark SF=10

Queries whose outputs are large

Has following-sibling Contains “//”

Our method is effective for

IO-intensive workloads

Natix showed better performance

for breadth-first traversals

(e.g., following-sibling)

21

Conclusions

� Proposed an efficient XML storage schemebase on DTM for iterative XQuery processing

� Our approach is effective for IO-intensive workloadssuch as queries including ‘//’.

� Document-ordered block allocation� Informed prefetching and scan-resistant caching

Summary

Future work� Automatic database tuning based on online analysis of data access patterns (e.g., buffer replacement policy and prefetching)

22

Thank you for your attention!

Questions?

23

Problem in XML-Relational mapping

<student><id>0551133</id><name>ichiro</name>

</student>

XML

Student

id name

XML Tree

A

B C

Hard disk

Relational Table

contentparentid

--A

0551133AB

ichiroAＣ

Mapping

24

-

-

-

R

0321CID

--9-7--46-NEXTSIB

1771123210PARENT

ETEETETEEETYPE

3

4

2

E

Updating facilities and versioning

When accessing to a record,

Logical address Physical address

For updating facilities, we change this method as follows:

Logical address Physical address

Address conversion table

25

Buffer management (XMark Q10 as an example)

080,673185.562Q

567,7021,919,586211.83LRU

buffer replacement

total read blocks

Elapsed time (msec)

14.2% speedups

20% of original document is returned

as the result

26

27

Rows

Blocks

Read in Pages

accessed row

Pages

dynamically

configured

nullRow Skeleton

Pyshical layout

28

DC/SD Normal TC/SD Normal

Access pattern analysis of XBench queries

29

#include <sys/mman.h>

void *mmap(void *start, size_t length, int prot, int flags,int fd, off_t offset);

int munmap(void *start, sizt_t length);

Memory Mapped DTM

We present a memory mapped scheme

extending the DTM model, it has boost

the performance significantly.