Collaborative Environments for Scientific Computing: The Task of - - PowerPoint PPT Presentation

▶
collaborative environments for scientific computing the
SMART_READER_LITE
LIVE PREVIEW

Collaborative Environments for Scientific Computing: The Task of - - PowerPoint PPT Presentation

Jan 03, 2023 152 likes •374 views

Collaborative Environments for Scientific Computing 11th ICMCM & SC 97 Collaborative Environments for Scientific Computing: The Task of Algorithm/Software Selection N. Ramakrishnan, A. Joshi, E.N. Houstis and J.R. Rice Department of

slide-1
SLIDE 1

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Collaborative Environments for Scientific Computing: The Task of Algorithm/Software Selection

  • N. Ramakrishnan, A. Joshi, E.N. Houstis and J.R. Rice

Department of Computer Sciences Purdue University, IN 47907

11th ICMCM & SC ‘97

slide-2
SLIDE 2

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Outline

  • PSEs for Scientific Computing
  • Advisory Systems in Scientific Computing
  • The PYTHIA Project
  • Networked Scientific Computing
  • Case Studies / Experimental Results
  • Conclusion

11th ICMCM & SC ‘97

slide-3
SLIDE 3

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Problem Solving Environments (PSEs)

  • What are PSEs?

“A PSE is a computer system that provides all the computational facilities necessary to solve a target class of problems.”

  • Features
  • Advanced Solution Methods
  • Automatic/SemiAutomatic Selection of Solution Methods
  • Ways to easily include new techniques
  • Other Features
  • Check formulation, select, assess correctness
  • Manage overall computational process
  • PSEs = User Interface + Libraries + Knowledge Bases + Integration
  • Example PSEs
  • MATLAB - Linear Algebra
  • Mathematica - Symbolic & Numeric Computation
  • ELLPACK, //ELLPACK, PDELab - PDE Problems
  • NetSolve - Optimization

11th ICMCM & SC ‘97

slide-4
SLIDE 4

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Advisory Systems in Scientific Computing

  • Necessary to supplement PSEs
  • Similar in idea to recommender systems
  • Automatic Algorithm Selection Systems
  • Motivations
  • Multiple Solution Approaches
  • Resource Limitations
  • Complex factors influencing applicability of algorithms
  • Computation Steering
  • Accept problem definition, performance criteria
  • Suggest software components and hardware resources
  • Example
  • “Use the 5-point star algorithm with a 200x200 grid.“
  • “Implement the algorithm on an nCUBE/2 with 16 processors.”

11th ICMCM & SC ‘97

slide-5
SLIDE 5

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

The PYTHIA Project

  • Automatic Algorithm Selection for
  • 2-D Linear, Elliptic PDEs
  • Numerical Quadrature
  • Uses
  • Databases of previously solved problems along with

performance information for various methods on them

  • Problem features for classifying problems
  • (Optimization) Group collections of problems with

“similar” features into problem classes

  • Mechanisms for comparing features to find “similar”

problems and problem classes

  • Three Key Components
  • PEM (Performance Evaluation Module)
  • AFD (Automatic Feature Determination)
  • KE (Knowledge Engine)

11th ICMCM & SC ‘97

slide-6
SLIDE 6

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Case Study 1: Numerical Quadrature

  • Uses
  • 124 routines
  • 286 test problems
  • Select a method to evaluate

an integral with a constraint

  • n relative accuracy so that

the number of function evaluations is minimized

  • Performance Evaluation
  • Run routines on applicable integrands and code results

as nfe predicates

  • Software “Exits”
  • ccurrence of roundoff error
  • difficulties in integrand behavior
  • divergent (or slowly) convergent integrals
  • limiting number of cycles

a b y = f(x) x y

11th ICMCM & SC ‘97

slide-7
SLIDE 7

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Case Study 1: Numerical Quadrature (Contd.)

  • AFD Module
  • Determines
  • whether integrand can be expressed as w(x)f(x)
  • the location of singularities of f
  • if there are singularities of f’
  • locations of discontinuities
  • range of integration
  • behavior of non-specific type
  • features of data points (on a grid)
  • User Input
  • Results
  • Best Algorithm Selection - 89%
  • Second Best Selection - 5%
  • ‘Reasonable’ Selection - 3%
  • ‘Hopeless’ Cases (!) - 3%

11th ICMCM & SC ‘97

slide-8
SLIDE 8

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Case Study 1: Numerical Quadrature (Contd.)

  • Rules uncovered

method (Problem, qags) :- accuracy(Problem, Acc), image(Problem, type6), sing(Problem), endptsing(Problem), noderivsing(Problem), range(Problem, finite).

  • Success mainly due to the methodology of data collection
  • Uncovered broad heuristics
  • similar to decision-tree models and lookup tables
  • confirmed to popular opinion
  • Examples
  • Adaptive algos. use lesser nfes than non-adaptive ones

(for high accuracy results)

  • Adaptive algos. take longer to meet low-accuracy constraints

11th ICMCM & SC ‘97

slide-9
SLIDE 9

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Case Study 2: PDEs

Classification Module Inference Engine User’s performance objectives: Error: less than 0.02% Memory Usage: less than 5 MB Execution Time: at most 10 minutes Service Charge: less than $10 PDE Problem Features:

  • perator, domain, boundary

conditions, I/O functions PYTHIA’s recommendation: Use 5-point star with a 200 x 200 grid on an nCube/2 using 16 processors. PYTHIA’s KB

11th ICMCM & SC ‘97

slide-10
SLIDE 10

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Networked Scientific Computing

  • A ‘Net-centric’ Scenario for Scientific Computing
  • Distributed Software + Information Resources
  • Need to be identified and ‘linked’ together
  • View of software changes from a product to a service
  • Improvements in QoS
  • Enabled by
  • HPCC technologies
  • Complexity of Application Software
  • Multi-disciplinary nature of large scale application research
  • Necessity to combine NC based supercomputer resources
  • Network Computational Servers
  • provide on-line access to PSEs
  • Need for Agent Based Environments

11th ICMCM & SC ‘97

slide-11
SLIDE 11

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Example Systems

  • Provide access to software libraries & documents
  • Netlib (UTK, ORNL, AT&T)
  • Enable PSEs to be accessed over the Web
  • Web //ELLPACK, Net //ELLPACK (Purdue)
  • NetSolve (UTK)
  • Index software modules acc. to functionality
  • GAMS (NIST)
  • Provide access to performance data, evaluation etc.
  • NEOS (ANL)
  • The Matrix Market (NIST)
  • Interface to HPCC repositories
  • NHSE (CRPC)

11th ICMCM & SC ‘97

slide-12
SLIDE 12

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Collaborative Scenario for Algorithm Selection

  • Assume
  • Multiple PYTHIA ‘servers’ each with specialized KB
  • Need
  • to model collaborative scenario and environments
  • to dynamically accommodate servers
  • to determine ‘most reasonable’ PYTHIA
  • to learn mappings from problems to PYTHIA servers
  • to handle (dis)appearance of servers with time
  • to provide ‘learning’ and ‘stable’ modes of operation
  • efficient mechanisms to switch between these modes

11th ICMCM & SC ‘97

slide-13
SLIDE 13

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

PYTHIA Agents

  • ‘Memory of Agents’
  • Two modes of operation
  • Learning Mode (LM)
  • Stable Mode (SM)
  • “Cache Invalidation” Strategies
  • Time Based
  • Reactive
  • Time based Reactive
  • Proactive
  • Learning facilitated by notion of reasonableness

Communication: KQML, TCP/IP Content: PYTHIA-TALK Local Data, Remote Agent Info. User Interface, Query Mechanism

11th ICMCM & SC ‘97

slide-14
SLIDE 14

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

PYTHIA Agents (Contd.) - Mechanisms for Collaboration

  • An Acceptance System for Agents
  • Based on Epistemic Utility Theory
  • Quantitative Measure of Reasonableness
  • based on Levi’s theory of knowledge
  • uses truth + utility
  • r(q) = p(q)Ut(q) + p(not(q))Uf(q)
  • p(q) determined by PYTHIA
  • Ut(q) = -Uf(q) = f(Ne) where Ne = number of exemplars
  • f: squashing function
  • A two-phase technique
  • an unsupervised technique to generate exemplars
  • a supervised method to learn mappings from problems to agents

11th ICMCM & SC ‘97

slide-15
SLIDE 15

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Overall Strategy

  • Establish PYTHIA servers with specialized KBs
  • Delegate a central PYTHIA agent
  • For each PDE problem
  • Query each agent
  • Determine most reasonable agent
  • Form one exemplar from this data (& learn)
  • Use the mapping so learnt
  • to determine agents for new problems
  • When does mapping become invalid?
  • when abilities of agents change (Dynamic Scenario)

11th ICMCM & SC ‘97

slide-16
SLIDE 16

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Experimental Results (Static Case)

  • Classes defined based on solution properties
  • Solution-Singular (6), Solution-Analytic (35), Solution-Oscillatory (34),

Solution-Boundary-Layer (32), Boundary-Conditions-Mixed (74)

  • These classes are not mutually exclusive!!
  • Multiple Agent Scenarios
  • 6,5,4 and 3 agents
  • 167 PDE problems
  • 111 problems - Larger Training Set
  • 56 problems - Smaller Training Set
  • Comparison with other learning mechanisms

11th ICMCM & SC ‘97

slide-17
SLIDE 17

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Experimental Results (Graphs)

  • Incremental learning of mappings
  • On-line adaptation
  • Gradual Reduction in Error

6 5 4 3

  • No. of Agents

10 20 30 40 50 60 70 80 90 100

% Accuracy

Our Technique RProp Plain BProp

Smaller Trg. Set

6 5 4 3

  • No. of Agents

10 20 30 40 50 60 70 80 90 100

% Accuracy

Our Technique RProp Plain BProp

Larger Trg. Set

11th ICMCM & SC ‘97

slide-18
SLIDE 18

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Experimental Results (Dynamic Scenario)

  • Time-Based Approach
  • Start with approx. 1/3 of KB in each PYTHIA agent
  • slowly infuse knowledge
  • Learn mapping at regular time intervals
  • Observations
  • Gradual improvement in accuracy to figures in static case
  • 40.85% to 98.20% in Scenario 1 (6 Agents)
  • 59.15% to 98.20% in Scenario 2 (5 Agents)
  • 59.28% to 98.20% in Scenario 3 (4 Agents)
  • 68.86% to 98.20% in Scenario 4 (3 Agents)

11th ICMCM & SC ‘97

slide-19
SLIDE 19

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Incremental Addition of Knowledge

10 20 30 40 50 60 10 20 30 40 50 60 70 80 90 100 Incremental Addition of Knowledge − Time Based Scheme Time Steps % Information about a class 11th ICMCM & SC ‘97

slide-20
SLIDE 20

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

More Graphs ...

10 20 30 40 50 60 30 40 50 60 70 80 90 100 Time Steps % Accuracy of PYTHIA−C Incremental Improvement in Accuracy − Time Based Scheme 11th ICMCM & SC ‘97

slide-21
SLIDE 21

Collaborative Environments for Scientific Computing Department of Computer Sciences, Purdue University

Conclusion

  • Collaborative Advisory Systems are feasible
  • Facilitate Networked Scientific Computing
  • Caters to a dynamic scenario of ‘servers’
  • Facility to learn on-line
  • Can switch from learning and stable modes of operation
  • Initial Results have been encouraging

For more details, pl. look at http://www.cs.purdue.edu/research/cse/pythia

11th ICMCM & SC ‘97