Efficient Processing of XML Path Queries Using the Disk-based F

1
Efficient Processing of XML Path Queries Using
the Disk-based FB Index

• Wei Wang
• University of New South Wales, Australia

With Hongzhi Wang (HIT), Hongjun Lu (HKUST),
Haifeng Jiang (IBM), Xuemin Lin (UNSW), Jianzhong
Li (HIT)
2
XML Query Processing

• XML
• Modeled as a labeled tree
• Query by structural constraint
• Simple Path Queries, e.g., //Customer//Name
• Branching/Twig Queries, e.g., //Customer//Zipcode
  //Name

3
Index or Join?
Q1 /a/b

• Index-based approaches
• DataGuide, 1-index
• FB Index
• and a few approximate indexes
• Join-based approaches
• Structural join
• Twig join

a
b
b
Join-based approaches appear to be more actively
researched!
4
Outline

• Introduction
• Disk-based FB Index
• Experiment
• Conclusions

5
XML Structural Indexes

• Exact Indexes
• 1-index
• Based on backward bisimilarity
• Covers all simple path queries
• FB Index
• Based on backward and forward bisimilarity
• Covers all branching queries (optimally)

6
A Running Example
Q1 /a/b
Q2 /a/bd
Q3 /a/bcd
7
Problems with FB Index?

• Lack of scalability
• Usually large in practice
• No immediate solution when it cannot be
  accommodated in memory
• Unbalanced, all-leaf-nodes tree
• Naïve solutions (e.g., B-tree, pre-order
  clustering in Lore, subtree clustering in Natix)
  do not work well
• Lack of efficiency
• Non-deterministic searching
• //-axis requires traversing the whole subtrees
• Much more costly when the index is not in the
  memory

8
Outline

• Introduction
• Disk-based FB Index
• Experiment
• Conclusions

9
Disk-based FB Index

• Overcome the memory limit by putting FB index to
  the disk
• Naïve method does not work well

Q1 /a/b
10
Basic Idea

• Moral Clustering is
• important
• Cluster by tag ? tape
• Cluster by parent ? segment block
• Cluster by 1-index ID ? chunk
• Benefits
• Optimized tree traversals
• Enable other intelligent algorithms

11
Q1 /a/b
12
Q.P. by Tree Traversal

• Dim 1 DFS/BFS
• Dim 2 Path/Branching Path
• Dim 3 / or //

Q5 /a/b/c
Q2 /a/bd
Q4 /a//c
Problem Still have to traverse the entire
subtrees to process //
13
Q.P. by RangeFetch

• H(1, c) 3, 6

(chunkID, tagName)
Q4 /a//c
Restriction Can only answer /p//q, where p is a
simple path.
14
More Data Structures

• 3 more tapes
• Add region code for each d-node in the extents ?
  Extents Tape
• Use physical (start, end) codes
• Sort d-nodes according to (start, end)
• Add Doc Tape
• Add Value Tape

15
Example
16
SegSJ

• Key observation
• Structural relationship between two segments can
  be inferred from the relationship between their
  first d-nodes in their extent.
• SegSJ(/p//q)
• R(s, e) ? A /p
• S(s, e) ? D //q
• Structural join R and S
• Using partition-based or sorting-based SJ
  algorithm

b1 ? (10,78), (210, 297),
d1 ? (19,25), (54, 66),
Take the (s, e) of the first d-node in each
segment
17
Outline

• Introduction
• Disk-based FB Index
• Experiment
• Conclusions

18
Experiments

• Setup
• DBLP/XMark/TreeBank
• 8 representative queries
• Dim 1 PC/AD
• Dim 2 Path/Twig
• Dim 3 Large/Small
• DFS, BFS, RangeFetch, SegSJ
• NoK, TwigStack, Kaushiks algorithm in SIGMOD
  04
• Metric time/PIO/LIO

19
Varying Buffer Size (PC-Path)
20
Varying Buffer Size (PC-Twig)
21
Varying Buffer Size (AD-Path)
22
Varying Buffer Size (AD-Twig)
23
Buffer Hit Ratio
24
Scalability
25
Comparing with Other Systems
26
Outline

• Introduction
• Disk-based FB Index
• Experiment
• Conclusions

27
Conclusions

• Disk-based FB Index
• Store and cluster the index on the disk
• More efficient and intelligent query processing
  algorithms
• Demonstrated good scalability and query
  efficiency
• Expecting new query processing algorithms based
  on index probing (in addition to join-based
  approaches)

28
QA

• Thank You!

29
Related Work

• Indexes
• Exact DataGuide, 1-index, FB Index
• Approx Approx. DataGuide, A(k)-index,
  D(k)-index, M(k)-index
• Join-based approaches
• Hybrid approach mixed-mode in VLDB 03
• Niagara
• VLDB 03 combines tree traversals joins
• SIGMOD 04 use 1-index to accelerate joins
• Clustering
• Lore pre-order
• Natix subtree