Transcript lcpc_xgli - Computer Science and Engineering

Compiler and Runtime Support for Shared
Memory Parallelization of Data Mining
Algorithms
Xiaogang Li
Ruoming Jin
Gagan Agrawal
Department of Computer and Information
Sciences
Ohio State University
Motivation

Languages, compilers, and runtime systems for highend computing


Can commercial applications benefit ?


Typically focus on scientific applications
A majority of top 500 parallel configurations are used as
database servers
Is there a role for parallel systems research ?


Parallel relational databases – probably not
Data mining, decision support – quite likely
Data Mining





Extracting useful models or patterns from large
datasets
Includes a variety of tasks - mining associations,
sequences, clustering data, building decision trees,
predictive models - several algorithms proposed for
each
Both compute and data intensive
Algorithms are well suited for parallel execution
High-level interfaces can be useful for application
development
Project Overview
Data Parallel Java
Compiler Techniques
FREERIDE(middleware)
Runtime Techniques
MPI+Posix Threads+File I/O
Clusters of SMPs
Outline





Key observation from mining algorithms
Parallelization Techniques
Middleware Support and Interface
Language Interface and Compilation techniques
Experimental Results



K- means
Apriori
Summary
Common Processing Structure


Structure of Common Data Mining Algorithms
{* Outer Sequential Loop *}
While () {
{ * Reduction Loop* }
Foreach (element e) {
(i,val) = process(e);
Reduc(i) = Reduc(i) op val;
}
}
Applies to major association mining, clustering
and decision tree construction algorithms
Outline





Key observation from mining algorithms
Parallelization Techniques
Middleware Support and Interface
Language Interface and Compilation techniques
Experimental Results



K- means
Apriori
Summary
Challenges in Parallelization


Statically partitioning the reduction object to avoid
race conditions is generally impossible.
Runtime preprocessing or scheduling also cannot be
applied



Can’t tell what you need to update w/o processing the
element
The size of reduction object means significant
memory overheads for replication
Locking and synchronization costs could be
significant because of the fine-grained updates to the
reduction object.
Parallelization Techniques




Full Replication: create a copy of the reduction object
for each thread
Full Locking: associate a lock with each element
Cache Sensitive Locking: one lock for all elements in
a cache block
Optimized Full Locking: put the element and
corresponding lock on the same cache block
Memory Layout for Various Locking Schemes
Full Locking
Cache-Sensitive Locking
Optimized Full Locking
Lock
Reduction Element
Outline





Key observation from mining algorithms
Parallelization Techniques
Middleware Support and Interface
Language Interface and Compilation techniques
Experimental Results



K- means
Apriori
Summary
Middleware Support for Shared Memory
Parallelization

Interface Requires:



Specification of an iterator and termination condition
Local reduction for each parallel loop
Functionality




Fetch data elements chunk by chunk, apply local reduction
Parallelization and Synchronization
Global reduction for all threads
Check termination condition, move to next iteration
Example :Kmeans Clustering Algorithm

Problem:
-Given N points in a metric space and a distance function.
-Try to find K centers and assign each point to one of these centers.
-Minimize total distance between each point and the center it belongs to.
• Algorithm
Make initial guesses for the centers m1, m2, ..., mk
Until there are no changes in any center
Use the estimated centers to classify the points into clusters
For i from 1 to k
Replace mi with the mean of all of the pointss for Cluster i
end_for
end_until
Programming Interface: k-means example

Initialization Function
void Kmeans::initialize() {
for (int i=0;i<k;i++) {
clusterID[I]=reducobject->alloc(ndim);
}
{* Initialize Centers *}
}
k-means example (contd.)

Local Reduction Function
void Kmeans::reduction(void *point) {
for (int i=0;i<k;i++) {
dis=distance(point,i);
if (dis<min) {
min=dis;
min_index=i;
}
for (int j=0;j<ndim;j++)
reductionobject->Add(objectID,j,point[j]);
reduction object->Add(objectID,ndim,1);
reductionobject->Add(objectID,ndim+1,dis);
}
}
Find a nearest center
Assign point to the
center
Outline




Key observation from mining algorithms
Middleware Support for Shared Memory
Parallelization
Interface and Compilation techniques
Experimental Results



K- means
Apriori
Summary
Language Support
A data parallel dialect of Java: to give compiler information
about independent collections of objects, parallel loops and
reduction operations
— domain & rectdomain
— foreach loop
— reduction interface:
can only be updated inside a foreach loop by operations
that are associative & commutative
-intermediate value of the reduction variables may
not be used within the loop, except for self-updates
-
K-means Clustering expressed by Data
Parallel Java
public class Kmeans {
public static void main(String[] args) {
RectDomain<1> InputDomain=[lowend:hiend];
KmPoint[1d] Input=new KmPoint[InputDomain];
While (not_converged) {
foreach (p in InputDomain) {
min=MIN_NUMBER;
for ( i=0;i<k;i++)
{ int dis=kcenter.distance(Input[p],i);
if(dis<min) {
min=dis;
minindex=i;
}}
kcenter.assign(Input[p],minindex,min);
}
kcenter.finalizing();
} }}
Input Data
Reduction Loop
Tasks of Compilation


Mapping from reduction interface in our dialet of Java
to reduction object used by middleware
- Parallelization techniques are transparent to compiler by
using reduction object.
Extract important function from Java code to fit
into our middleware
-Data fetching
-Local reduction
-Iterator and termination condition
Mapping of Reduction interface

Decide the size of reduction object to be allocated.
-By declaration information of reduction interface
-By symbolic analysis if can not decide statically

Allocation of reduction object
-Layout can be block or cyclic

Changed reference and modification of members to
corresponding elements of reduction object.
x[1]=0  (*reductionElement)(reduct_buffer,1)=0
Extract important functions

Local reduction function
-From body of data parallel loop
-Cumulative and associative operations on reduction interface are
replaced by operator of reduction object.
meansx1[i]+=Input[p].x2
 reducObject->Add(reduct_buffer, I,Input.x1)

Iterator and termination
-simple from overall code

Data fetching function
- from declaration of input class.
-use constructor of input class to provide additional information.
Results


Full Replication
achieveoutperforms
best result when
Cache
Sensitive locking
Full
size of reduction
object isFull
small
replication
and Optimized
locking as
size of reduction object increased
Relative performance of Full Replication, Optimized Full locking and CacheSensitive Locking : 4 threads, different support levels
Results
Comparison of compiler generated and manual versions– Apriori Association
Mining (1GB Dataset)
Results
Comparison of compiler generated and manual versions– K-means
Clustering ( 1GB Dataset, K=100)
Conclusion
Provide runtime and compiler supports for shared
parallelization of data mining applications.
-Different parallelization techniques.
-Support of middleware simplifies code generation.
-Compiler generated code is competitive.