Architecture Conscious Data

Mining
Srinivasan Parthasarathy
Data Mining Research Lab
Ohio State University
 KDD & Next Generation
Challenges
• KDD is an iterative and interactive process the goal of
which is to extract interesting and actionable
information from potentially large data stores efficiently
• Young field, long laundry list of technical challenges
– Theoretical foundations in various sub-fields
– Interestingness and Ranking
– New and Exciting Applications
• Embedding domain knowledge effectively
– Visualization for data & model understanding
– Efficient and scalable algorithms (focus of this talk)
• Other challenges
– Educational (talk a bit about this at the end)
– Reproducability (need for benchmarks)
– Socio-Political
 Efficiency in the KDD process
• Why is it important?
– Interactive nature of KDD
– Real-time constraints
• What makes it challenging?
– Dataset properties (large,
heterogeneous, distributed)
– Computational complexity Mining Simulation Data
• Example Applications
– Clinical data Diagnosing disease and
modeling progression
– Biological data Twa et al 2005
Analyzing (dynamic) Networks
– Large scale simulation data Protein Interaction Network (yeast)
– Social network data
– Sensor data, WWW data….
 Toward Efficient Realizations
• Data driven approach
– Compression, Sampling, Dimensionality Reduction, Feature
Selection, Matrix Factorization etc.
• Computational driven approach
– Intelligent search space pruning to reduce complexity
– Approximate algorithms, streaming algorithms
– Parallel and distributed algorithms
• Architecture-Conscious approach (this talk)
– Largely orthogonal to the above alternatives
– Objective is to understand limitations and novel features of
modern and emerging architecture(s)
– Subsequently, re-architect algorithms to better utilize system
resources.
 Houston, do we have a problem?
• Turns out we do
– Many state-of-the-art data mining algorithms grossly
under-utilize processor resources [Ghoting 2005]
• Why?
1. Data intensive algorithms – lots of memory accesses
– high latency penalty.
2. Mining algorithms are extremely irregular in nature –
data and parameter driven – hard to predict
3. Use of pointer-based data structures – poor ILP
4. Do not leverage important features of modern
architectures – automated compiler/runtime systems
are handicapped because of 1, 2 and 3.
 Spatial Locality
• Improve spatial locality of r
dynamic data structures
a f c
– Memory pooling
– Loss-less compression – c b b
store only data that is
needed – allows for more f p
data per cache line
– Memory placement to m b

match dominant access

order p m

– Side benefit – enables

effective hardware
prefetching (latency
alleviating mechanism) DFS allocation
 Temporal Locality and Leveraging
SMT r

• Data Structure Tiling

– Operate on a tile-by-tile
basis
• Non-overlapping
(traditional)
• Overlapping
• Smart data partitioning
– Jigsaw puzzle analogy
• SMT
– Co-schedule tasks that
operate on same data tile
helps improve
performance Tile 1 Tile 2 Tile N-1 Tile N
 Sample Benefits
• Gains in performance can be
staggering
– Frequent patterns (itemsets,
trees, graphs)
– Outlier Detection
– Clustering
• Benefits to end applications
– Scientific simulation data
– Web data
– Molecular and Clinical data
• For network of workstations
– minimize communication and VLDB’05, KDD’06, VLDBJ07
leverage remote memory PPOPP’07
– Enables mining of terabyte scale
distributed datasets efficiently.
 CMPs (next frontier)
• Why the push from • Challenges
industry? – Existing applications, they
– Increasing clock need to be rewritten to use
frequencies is not returning multiple threads of
improved IPC, and it is execution
increasing power costs and – Compiler and runtime
thermal issues techniques have a hard
• Two new PCs in my den, time already – application
must help
no need for the heat vent! – Fine-grained sharing of
– Great for winters! processor resources
• Importantly (cache, bus/channel etc.)
– Parallel Computing meets – Memory hierarchy issues
mainstream commodity are even more challenging
market • Potential solution
– Adaptable algorithms
 Adaptive algorithms
• Key idea:Trading off • Key idea: Moldable
memory for redundant partitioning and adaptive
computation scheduling of tasks
• Benefits: • Benefits
– Reduced working set sizes – Better CPU utilization
– Likely to have reduced – If co-scheduling – reduced
bandwidth pressure cache miss rates
– Utilizing strengths of the • Challenges:
CMP – Sensing the problem
• Challenge: – Re-architecting algorithm
– Sensing the problem • Moldable task
– Re-architecting algorithm to decomposition
reduce memory • Pass on enough state to
consumption move task to another core
 Adaptive algorithms performance
• Graph mining • Tree Mining
– Converted to sequence space
– Gaston vs. Gspan vs. (dynamic arrays)
Hybrid (adaptive) • Better locality, ILP
– Reduced memory LCS
matching + structure checks
– Leveraged hybrid scheduling
– Sequential Performance
• 2 order reduction in
memory footprint
• 3 orders improvement in
processing time
– Parallel Performance
• Linear scalability on a 4-core
dual chip (8 cores)
• Adapted similar idea to XML
indexing with similar results!

ICDM’06, CIKM’06, VLDB’07

 Esoteric CMPs (CELL)
• Interesting design point on 1000
commodity CMP space
– 25 GB/s OC bandwidth 100
– 8 cores (SPUs) + 1 PPU kMeans
– FP computation 200 GFlops 10 KNN

– Orca
Breakthroughs in commodity
1
processing
• Challenges

on

8
2

D2
0

l l-

l l-
Xe
0.1

25
um

Ce

Ce
m
– Hard to program

ni

iu
ro
It a

nt
te

Pe
op
– Need to explicitly manage
memory and data transfers
between PPU and SPUs Cell-6 on Sony Playstation
– Probably not suitable for all Cell-8 is simulated
programs
All cases codes optimized and
– Interesting class of algorithms
and kernels can benefit Implemented on appropriate compiler
significantly!
 Mining on Clusters
• Heavily researched over the last 15 years
– DDM Wiki (a very nice start point resource)
• What are the “new” challenges?
– Non-homogeneous “hybrid” clusters – (e.g. Roadrunner)
– Multi-level parallelism (on chip, on node, on cluster)
– Leveraging features of high end systems networking
• Infiniband makes it feasible and cheaper to access remote memory
than local disk – how to leverage?
– KDD may be particularly amenable to pipelined parallelism – a
largely ignored approach
– KDD and the grid (heard about this yesterday)
– Application specific challenges -- e.g. astronomy, folding@home
etc.
 Discussion
• KDD is an iterative and interactive process the goal of
which is to extract interesting and actionable information
from potentially large data stores efficiently
• This talk was primarily about the last but all 3 are
important.
• Architecture conscious data mining is a viable orthogonal
approach to achieve efficiency (references in paper)
– Tangible benefits to applications, algorithms and kernels
– Lower memory footprints + significantly faster performance
– Adaptive algorithms are necessary for emerging architectures
– Whats next? Services oriented architecture
• Plug-and-Play naturally connects with KDD process
• An effective mechanism to keep cores busy.
 Broadly Speaking
• Education
– As an aside parallel algorithms and high performance
computing has to be a part of basic CS curriculum.
– We as data-intensive science need to understand the
key systems issues better from OS and architecture
friends
• Broader Scientific Impact
– Interactions between Systems and Data Mining
• Data mining for software engineering, invariant tracking,
testing, bug detection in sequential and parallel codes
• Data mining for performance modeling
• Leveraging systems features for data mining
 Thanks
• Students
– A. Ghoting, G. Buehrer, S. Tatikonda
• Collaborating Colleagues
– OSU-Physics, OSU-Biomedical Informatics, Intel, IBM
• Funding agencies
– NSF CCF0702587, CNS-0406386, CAREER-IIS-0347662, RI-
CNS-0403342.
– DOE Early career principal investigator grant
– IBM Faculty partnership
• Organizers of this workshop
• Additional Information: dmrl.cse.ohio-state.edu or
srini@cse.ohio-state.edu