Upload
jillian-erickson
View
20
Download
2
Embed Size (px)
DESCRIPTION
ALPS: Software Framework for Scheduling Parallel Computations with Application to Parallel Space-Time Adaptive Processing Kyusoon Lee and Adam W. Bojanczyk Cornell University Ithaca, NY, 14850 {KL224,AWB8}@cornell.edu. High Performance Embedded Computing (HPEC) Workshop 18 – 20 September 2007. - PowerPoint PPT Presentation
Citation preview
ALPS: Software Framework for Scheduling Parallel Computationswith Application to Parallel Space-Time Adaptive Processing
Kyusoon Lee and Adam W. BojanczykCornell UniversityIthaca, NY, 14850
{KL224,AWB8}@cornell.edu
High Performance Embedded Computing (HPEC) Workshop18 – 20 September 2007
/* This is a .cpp file of the partially adaptive STAP methods that the user had in mind.
This is optimized for minimum latency or maximum throughput on a given parallel machine.*/
#include <alps.h>...
Various STAP heuristicsconstructed from ALPS building blocks
Modeling execution times ofbasic building blocks under various configurations
Find optimal configuration that minimizesoverall makespan or maximizes throughput
ALPS CodeGenUser’s idea on
Parallel Computations(e.g. STAP)
Task Graph
ALPS Benchmarker & Modeler
ALPS Scheduler
ALPS Library
ALPS Software Framework
Building execution time models using incremental weighted least-squares
Measurements are not quite reliable
Number of variables in the model is largeDifficulties
0 0.1 0.2 0.3 0.4 0.5 0.6 0.70
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Relative errors from ordinary least-square
Relative errors from weighted least-square
10% be
tter
20% be
tter
30% be
tter
40% be
tter
7.5% errorfrom INV-weightedleast-square
26.6% errorfrom ordinaryleast-square
Scheduling tree-shaped task graphs on parallel computers exploiting mixed-parallelism
Communication cost not negligible
Arbitrary computation costs
Trade-off between task-parallelism and data-parallelism
Difficulties
0
10
20
30
Data-parallelismonly
Task-parallelismpreferred
Ourscheduling algorithm
Makespan (sec)
estimated achieved
0
0.05
0.1
Data-parallelismonly
Task-parallelismpreferred
Ourscheduling algorithm
Throughput (1/sec)
estimated
achieved