1. LCPC02 Wei Du Renato Ferreira Gagan Agrawal
Towards Compiler Support for Data Intensive Applications on Distributed Heterogeneous Resources
Wei Du
Renato Ferreira
Gagan Agrawal
Ohio-State University

2. Ohio-State University
Motivation
Grid Environment
geographically distributed heterogeneous resources
Scientific and commercial data intensive applications
generalized reduction operations are very common in the processing structure
No compiler support for high-level languages for grid application development
7/26/02
Ohio-State University

3. Software Architecture
Data Parallel Java
assume all data are available in a flat memory
assume all computation are done on a single processor
extensions of Java:
domain & rectdomain
foreach loop
reduction variables
Data Parallel Java
Compiler Support
Filter-stream Program
On DataCutter
7/26/02
Ohio-State University

4. Example code public class kNN { static buffer kbuffer;
public static void main(String[] args) {
double dis;
Point<3> lowend =[0,0,0];
Point<3> hiend =[Integer.parseInt(args[0]),
Integer.parseInt(args[1]),
Integer.parseInt(args[2])];
Point<3> p;
RectDomain<3> InputDomain=[lowend:hiend];
kPoint[3d] Input=new kPoint[InputDomain];
foreach (p in InputDomain) {
if (Input[p].inRange(R))
dis=Input[p].distance(W);
kbuffer.insert(Input[p],dis); }

5. DataCutter — Grid development tool
ongoing project at University of Maryland / OSU ( Beynon, Kurc, Sussman, Saltz et al.)
targets distributed, heterogeneous environments
decomposes application-specific data processing operations into a set of interacting processes
provides a set of interfaces
filter
stream
layout & placement
stream1
stream2
filter1
filter2
filter3

6. Compiler Overview Data Parallel Java Data Centric Transformation
Loop Fission
Global
Reduction
Analysis
Filter Enumeration
Granularity
Selection
Final Code
Generation
Filter-Stream programming
7/26/02
Ohio-State University

7. Experience with a Mining Algo.
K-Nearest Neighbors
Given a 3-D range
R= <(x1, y1, z1), (x2, y2, z2)>,
and a point  = (a, b, c).
We want to find the nearest K neighbors of  within R.
Range_Query
Input[p].inRange?
discard
Local reduction
Dis = … …
Kbuffer.insert(…, …)
Range_query
Select
Combine
R-S stream
S-C stream
Read
data
7/26/02
Ohio-State University

8. Experimental Results Experimented on a LAN
Data is available on 2 machines
Results are on the third machine
Dataset contains 3-D points, of size 1.2M and 12M
K is 20
Consider 0, 1, 4, 16 other computation intensive jobs on machines hosting data
Running time is in milliseconds
Performance difference of the two versions is within 15%
Comparing Manual Version and compiler-generated Version

9. Experience with Image Querying
Virtual Microscope
Input:
a digitized image,
a rectangular region R,
a subsampling factor
Output:
an enlarged image for the specified image region with a particular sampling factor
querybox = [lowend, highend]
foreach (p in querybox) {
if (VScope[p].is_sampled(lowend,
subsampling_factor)
q = (p-lowend)/subsampling_factor;
Output[q].Assign(VScope[p]); }
7/26/02
Ohio-State University

10. Experimental Results Experimented on a LAN
Data is available on machine other than where the results are to be displayed
Image is of size 800k
Sampling factor are 4 and 16
Consider 0, 1, 4, 16 other computation intensive jobs on machines hosting data
Running time is in milliseconds
Performance difference of the two versions is within 15%
Comparing Manual Version and compiler-generated Version

11. Summary and Future Work
Aims at developing compiler support for using heterogeneous and distributed resources for processing geographically distributed datasets
Experimented on simple data intensive codes with simple filter generation heuristics – Initial results are quite encouraging
Future Work:
more applications (visualization)
more compiler analysis (sophisticated heuristics, loop fission, data-centric code generation … …)
7/26/02
Ohio-State University

12. Thank you !!!

13. Compiler Overview Data Parallel Java Data Centric Transformation
Loop Fission
Global
Reduction
Analysis
Filter Enumeration
Granularity
Selection
Final Code
Generation
Filter-Stream programming
10/17/2019
Ohio-State University

14. Ohio-State University
Data Parallel Java
· assume all data are available in a flat memory
· assume all computation are done on a single processor
· extensions of Java: to give compiler information about independent
collections of objects, parallel loops and reduction operations
— domain & rectdomain
— foreach loop
— reduction variables:
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
10/17/2019
Ohio-State University

15. Ohio-State University
Why DataCutter ?
Typical DDM algorithm
DataCutter features
local reduction on geographically dispersed data
global reduction to combine the results
decomposition of application into a set of filters
filters are location independent
filters interact with each other via streams
10/17/2019
Ohio-State University

16. K-nearest neighbor search algorithm on DataCutter — A case study
Problem definition:
Given a 3-D range
R= <(x1, y1, z1), (x2, y2, z2)>,
and a point  = (a, b, c).
We want to find the nearest K neighbors of  within R.
Solution:
Range_query
Select
Combine
R-S stream
S-C stream
Read
data
10/17/2019
Ohio-State University