Inside an XSLT Processor

1
Inside an XSLT Processor

• Michael Kay, ICL
• 19 May 2000

2
About me

• ICL Fellow, systems architect
• Database background
• Developer of SAXON
• Author of XSLT Programmers Referencepublished
  by Wrox Press
• Recently joined XSL WG as invited expert

3
About this talk

• The XSLT Processing Model
• Structure of an XSLT Processor
• Performance
• current limitations
• possible ways forward
• Ideas on future development of the language

4
The XSLT Processing Modelfirst approximation
Style
sheet
Source
Result
Document
Document
Transformation
Process
5
The XSLT Processing Modelin more detail
Style
sheet
Parsing
Serialization
Stylesheet
Tree
Source
Result
Document
Document
SourceTree
ResultTree
TransformationProcess
6
An XSLT Template Rule
Pattern
ltxsltemplate match"appendix/para1"gt
lth4gt ltxslnumber level"single"/gt
ltxslvalue-of select"_at_title"/gt lt/h4gt
ltpgt ltxslapply-templates/gt
lt/pgt lt/xsltemplategt
ResultElement
XPathExpression
Instruction
7
Architecture of an XSLT processor
XML Parser
Stylesheet
Tree Builder
XPathcompiler
XSLTcompiler
Compiled Stylesheet
XSLT interpreter
XPath interpreter
Source
Result
XMLParser
TreeBuilder
OutputManager
XML serializer
HTML serializer
Text serializer
SourceTree
8
At compile time

• Parse and validate the stylesheet
• Parse and validate all XPath expressions
• and attribute value templates
• Build rule base for matching patterns
• Resolve references to named variables, functions,
  and templates
• Flatten the import tree
• Optimize XPath expressions

9
Where does the time go?
Serialize
Build Source
Output
Tree
Process
Templates
Compile
Stylesheet
10
Is Performance a Problem?

• Client side usually not
• XSLT processing is generally faster than download
  speed
• Server side sometimes
• CPU usage when handling very high throughput
• Memory problems when handling very large
  documents

11
Some performance tips

• Keep documents small split them first
• Process once, at publishing time
• or use caching
• Do several simple transforms in series
• Avoid complex patterns in template rules
• Use keys
• Use external functions
• Avoid "//item"

12
Performance progress
Simpleoptimization
Stylesheet compilation Java code
optimization Lazy evaluation Simple XPath
optimization Tail recursion
20 sec/Mb
Incremental parsing Pipelining Use of
schema Pattern matching Full XPath
optimization Compile to bytecodes
Advancedoptimization
5 sec/Mb
1 sec/Mb
Today
13
Interesting research areas

• Database integration transforming a document
  without loading into memory
• Applying regular expression theory
• Execution as a sequence of serial passes
• Using schema knowledge at compile time
• Eager node numbering

14
Potential language features

• Serial transformation language?
• Multi-pass stylesheets
• Higher-level "relational" constructs grouping,
  joins, logical quantifiers
• Richer data types
• Assignment statement ????

15
Summary

• XSLT language is now stable
• XSLT processor technology is starting to be well
  understood
• First crop of products are capable of significant
  performance
• Now the research needs to start on the next phase
  of optimization techniques