Issues in Parallelizing Pattern Mining Workloads

1
Issues in Parallelizing Pattern Mining Workloads

• Shirish Tatikonda
• CSE 788Z11

05th December, 2006
2
Outline

• Pattern mining algorithms
• Parallelization
• Issues
• Possible solutions

3
Pattern Mining Algorithms

• Patterns
• Transactions
• Sequences
• Trees
• Graphs
• Goal
• Given a large database of transactions,
• a threshold s
• Need the of set of all frequent patterns
• Frequent of occurrences s

Walmart Data T1 Milk, Bread, Beer, Diaper T2
Milk, Beer T3 Milk, Bread, Ice cream T4 Bread,
Beer, Diaper s 3 Frequent Milk, Bread,
Beer, Bread,Beer
4
Generic Approach

• Candidate Generation
• Level-wise approach
• Pattern growth approach
• Support Counting
• Evaluate each candidate

5
Pattern growth approach

• Depth-first traversal
• Equivalence classes
• Seed pattern
• Growing method
• Apriori Principle
• For pruning

Search Space Traversal
Infrequent
Seed
Frequent
Pruned
6
Parallelization

• Coarse-grain Equivalence Class Level
• Load balancing
• Fine-grain Each Pattern Level
• High overhead

P1
P2
P3
P4
P5
P6
Parallelizing equivalence classes
7
Issues

• Considering only the shared-memory systems
• Load imbalance
• In case of coarse-grain parallelization
• Speculative Execution
• Large working sets
• In case of CMPs
• Algorithmic rather than architectural solutions

8
Related Work

• Thread-Level Speculation (TLS)
• Mitosis compiler PLDI 2005
• Speculative thread level parallelism (SpMT)
• Speculative threads
• Executes a part of the original application
• When finished, speculation is verified
• Determines the spawning points
• Loops, basic blocks, subroutines etc.
• Inter-thread dependencies
• Synchronization
• Value prediction using pre-computation(p-) slices
• Compute p-slices using speculative optimizations

9
Related Work

• POSH compiler PPOPP 2006
• Fully automated TLS compiler
• Three phases
• Task selection
• based on programs substructures
• value prediction
• Spawn hoisting
• Task refinement
• to improve the quality of chosen tasks
• through profiling

Parallelism
Pre-fetching
10
Speculative Execution (1)
P2
P1
P3
11
Speculative Execution (2)

• Performance improvement
• High
• Marginal
• None
• Negative
• Depends on dataset characteristics
• Characterization
• Application in data mining
• Speculatively execute future iterations
• Tracking followed by classification

12
Large Working Sets

• Support Counting
• Evaluate a candidate pattern
• Scan a subset of database transactions
• Working set can potentially be large
• CMP (vs SMP)
• Less cache memory available to each processor
• Less main memory shared by all processors

Working set has to be reduced to
limit the bandwidth utilization
13
(No Transcript)