Implementing Scalable Tree-based Algorithms using MRNet

1
Implementing Scalable Tree-based Algorithms
using MRNet

• Ting ChenMark Cowlishaw

2
Introduction

• Background on MRNet
• Our Applications
• Reverse index
• Online queries
• Description of Experiments
• Progress
• Next steps

3
Background MRNet Roth, Arnold, Miller03

• Tree-Based Overlay Network
• Nodes of a distributed application are arranged
  in a tree-structure
• Leaves producing data that is aggregated and
  filtered by higher levels of the tree
• Separation of programs and their running tree
  topology
• A TBON program can run on tree-network of
  different topologies

4
Reverse indexes for Keyword Queries

• A keyword query is a list of words lt w1,w2,
  ...,wn gt .
• A typical reverse index is a list of words
• Each word points to a list of document IDs
  containing the word.
• A document ID list can either be sorted by
  document names or by the number of times the word
  appears in the document.

Wisconsin
Badgers
DocId No. of Occurences
B.htm C.htm . 6 4
DocId No. of Occurences
A.htm B.htm C.htm . 2 7 5
5
On-line queries N-best frequency (N2)
DocC 7
DocB 3
DocD 0
DocF 6
DocA 5
DocD 0
DocC 7
DocF 6
DocB 3
DocA 5
DocE 2
6
Objective / Experiments

• How does tree topology affect application
  performance
• Macro-benchmarks throughput/time-to-completion
  for index building and response time for on-line
  queries
• Micro-benchmarks the amount of total IO, I/O
  performed for each node, data transferred
• Scale-Up and Speed-Up curves with the increase of
  cluster nodes
• Is Multiple-level ( gt 2) helpful?

7
Progress - data

• Experiment Data
• All Wikipedia documents
• 8GB of data
• 4 Million documents
• Probably enough for a tree with 32 leaves
• Data in Wiki-text format

8
Progress Design / Implementation

• Message Formats
• Document/keyword delivery
• Acknowledgment
• keys / statistics
• Messages for Microbenchmarks
• Shared Classes
• DocumentStatistics (back end) in test
• StatisticsEntry (all) in test
• StatisticsList (all)
• Familiarity with MRNet Toolset
• Debugging MRNet Programs

9
Next Steps

• Deploy at medium scale (32 Nodes)
• Experiments to determine fan-out
• Maximize throughput (bytes/second)
• Minimize time-to-completion
• Microbenchmarks
• Collected and filtered using MRNet
• Idle time
• Messages per second, total message traffic

10
Futures

• Replace N-best with distance from median
• Add support for new document types
• XML
• HTML
• Generalize to other MapReduceDean04
  Applications
• More realistic relevance ranking
• Reverse hyperlink count
• Collection term vectors