Presentation is loading. Please wait.

Presentation is loading. Please wait.

About the Capability of Some Parallel Program Metric Prediction Using Neural Network Approach Vera Yu. Goritskaya Nina N. Popova

Published byRoss Walsh Modified over 8 years ago

Similar presentations

Presentation on theme: "About the Capability of Some Parallel Program Metric Prediction Using Neural Network Approach Vera Yu. Goritskaya Nina N. Popova"— Presentation transcript:

1 About the Capability of Some Parallel Program Metric Prediction Using Neural Network Approach Vera Yu. Goritskaya vera@angel.cs.msu.su Nina N. Popova popova@cs.msu.su

2 Problem Area Rising complexity of multiprocessor systems Heterogeneous clusters Distributed systems which include clusters and other multiprocessors Parallel program specifics Parallel program execution is affected by many factors communication environment loading nodes on which application is scheduled can vary … Almost impossible to estimate how a program would behave when running on multiprocessor

3 Parallel Program Metric Examples Run time (flow time) Tflow = Tcomputation + Tcommunication + Tidle Speedup = T1 / TN Efficiency = Speedup / N Scalability of a parallel system Efficiency is the same for increasing the number of processors and the size of the problem …

4 Related Works Neural network mechanism for application performance prediction Ipek, B.R. de Supinski, M. Schultz, S.A. McKee “An Approach to Performance Prediction for Parallel Applications” // Euro-Par 2005 Parallel Processing, Volume 3648, 2005, p. 196-205, ISBN 978-3-540-28700-1, 2005 Analytic modeling for performance tuning of parallel programs Crovella Mark E., Tomas J. LeBlanc. Parallel Performance Prediction Using Lost Cycles Analysis // Proceedings of Supercomputing’94, 1994. P. 600-610 Job execution time estimation based on program source analysis V.V. Balashov, A.P. Kapitonova, V.A. Kostenko, R.L. Smelyanskiy, N.V. Yuschenko “Method for estimating platform-optimized application execution time based on its high-level language source code” // Proceedings of 1 st international conference “Digital signal processing and its applications”, Volume IV, p. 203-220

5 Project Features Parallel application flow time prediction i.e. the time that program would spend inside the multiprocessor system Consider large amount of parameters communication environment loading nodes characteristics scheduling features … Neural network mechanism potentially can be applied to various parallel systems and applications

6 Project Features (2) Processes data: can be gathered without affecting source code of an application can be gathered using standard OS and job managing system utilities includes: job submission moment required processors maximum required execution time system loading at the submission moment size of executable, etc.

7 Project Features (3) Improving prediction accuracy: we gather characteristics of multiple executions of an application (“execution history”) sample “historical” characteristics: average execution time for definite application with fixed required processors number average required time for definite application average size of executable, etc.

8 Data Pre-Processing Grouping parallel programs from input set into categories according to the average execution time 4 groups in described case ( 5000 sec) “Noise” data excluding samples with max and min execution time values for each job were removed from input data sets samples corresponding to rejected jobs were also excluded

9 Target Platforms IBM eServer pSeries 690 (“Regatta”) 16-processor SMP architecture IBM eServer pSeries 360 (“Hill”) 10-processor cluster

10 Neural Network Architectures Multilayer feedforward network with sigmoid transfer function (1 hidden layer) Elman backpropagation network

11 Training Results Job group (according to the average execution time) Performance (multiplayer feedforward network with sigmoid transfer function) Performance (elman backpropagation network) < 100 sec10 -2 – 10 -3 10 -3 100 – 1000 sec10 -4 1000 – 5000 sec10 -4 – 10 -5 10 -4 > 5000 sec10 -5

12 Testing: Execution Time Prediction (“Regatta”) Execution time prediction (feedforward NN) on “Regatta” Execution time prediction (Elman NN) on “Regatta” predicted values real values

13 Target Platforms IBM eServer pSeries 690 (“Regatta”) 16-processor SMP architecture IBM eServer pSeries 360 (“Hill”) 10-processor cluster

14 Neural Network Architecture Elman backpropagation network

15 Testing: Execution Time Prediction (“Hill”) predicted values real values tasks time (sec) Hill’s flow is more homogeneous than Regatta’s flow

16 Conclusions and Future Work Improving data processing methods possibly will lead to more accurate results Using described approach new scheduling algorithms can be developed Applying NN prediction mechanisms on other multiprocessor platforms Problem Solving Environments

17 Thank you! Q & A ?

Download ppt "About the Capability of Some Parallel Program Metric Prediction Using Neural Network Approach Vera Yu. Goritskaya Nina N. Popova"

Similar presentations

1 Lecture 5: Part 1 Performance Laws: Speedup and Scalability.

1 Lecture 5: Part 1 Performance Laws: Speedup and Scalability.
History of Distributed Systems Joseph Cordina

History of Distributed Systems Joseph Cordina
A Parallel Computational Model for Heterogeneous Clusters Jose Luis Bosque, Luis Pastor, IEEE TRASACTION ON PARALLEL AND DISTRIBUTED SYSTEM, VOL. 17, NO.

A Parallel Computational Model for Heterogeneous Clusters Jose Luis Bosque, Luis Pastor, IEEE TRASACTION ON PARALLEL AND DISTRIBUTED SYSTEM, VOL. 17, NO.
Neural Networks. R & G Chapter 8 8.1 Feed-Forward Neural Networks otherwise known as The Multi-layer Perceptron or The Back-Propagation Neural Network.

Neural Networks. R & G Chapter Feed-Forward Neural Networks otherwise known as The Multi-layer Perceptron or The Back-Propagation Neural Network.
Fault-tolerant Adaptive Divisible Load Scheduling Xuan Lin, Sumanth J. V. Acknowledge: a few slides of DLT are from Thomas Robertazzi ’ s presentation.

Fault-tolerant Adaptive Divisible Load Scheduling Xuan Lin, Sumanth J. V. Acknowledge: a few slides of DLT are from Thomas Robertazzi ’ s presentation.
1 Introduction to Load Balancing: l Definition of Distributed systems. Collection of independent loosely coupled computing resources. l Load Balancing.

1 Introduction to Load Balancing: l Definition of Distributed systems. Collection of independent loosely coupled computing resources. l Load Balancing.
Classification and Prediction by Yen-Hsien Lee Department of Information Management College of Management National Sun Yat-Sen University March 4, 2003.

Classification and Prediction by Yen-Hsien Lee Department of Information Management College of Management National Sun Yat-Sen University March 4, 2003.
1 Neural plug-in motor coil thermal modeling Mo-Yuen Chow; Tipsuwan Y Industrial Electronics Society, 3000. IECON 26th Annual Conference of the IEEE, Volume:

1 Neural plug-in motor coil thermal modeling Mo-Yuen Chow; Tipsuwan Y Industrial Electronics Society, IECON 26th Annual Conference of the IEEE, Volume:
Neural networks - Lecture 111 Recurrent neural networks (II) Time series processing –Networks with delayed input layer –Elman network Cellular networks.

Neural networks - Lecture 111 Recurrent neural networks (II) Time series processing –Networks with delayed input layer –Elman network Cellular networks.
PARUS: a parallel programming framework for heterogeneous multiprocessor systems Alexey N. Salnikov (salnikov cs.msu.su) Moscow State University Faculty.

PARUS: a parallel programming framework for heterogeneous multiprocessor systems Alexey N. Salnikov (salnikov cs.msu.su) Moscow State University Faculty.
Computer System Architectures Computer System Software

Computer System Architectures Computer System Software
Development in hardware – Why? Option: array of custom processing nodes Step 1: analyze the application and extract the component tasks Step 2: design.

Development in hardware – Why? Option: array of custom processing nodes Step 1: analyze the application and extract the component tasks Step 2: design.
Soft Computing Colloquium 2 Selection of neural network, Hybrid neural networks.

Soft Computing Colloquium 2 Selection of neural network, Hybrid neural networks.
Cmpe 589 Spring 2008. Software Quality Metrics Product  product attributes –Size, complexity, design features, performance, quality level Process  Used.

Cmpe 589 Spring Software Quality Metrics Product  product attributes –Size, complexity, design features, performance, quality level Process  Used.
Predictive Runtime Code Scheduling for Heterogeneous Architectures 1.

Predictive Runtime Code Scheduling for Heterogeneous Architectures 1.
 What is an operating system? What is an operating system?  Where does the OS fit in? Where does the OS fit in?  Services provided by an OS Services.

 What is an operating system? What is an operating system?  Where does the OS fit in? Where does the OS fit in?  Services provided by an OS Services.
1 Interconnects Shared address space and message passing computers can be constructed by connecting processors and memory unit using a variety of interconnection.

1 Interconnects Shared address space and message passing computers can be constructed by connecting processors and memory unit using a variety of interconnection.
Multiple-Layer Networks and Backpropagation Algorithms

Multiple-Layer Networks and Backpropagation Algorithms
Artificial Neural Nets and AI Connectionism Sub symbolic reasoning.

Artificial Neural Nets and AI Connectionism Sub symbolic reasoning.
IE 585 Introduction to Neural Networks. 2 Modeling Continuum Unarticulated Wisdom Articulated Qualitative Models Theoretic (First Principles) Models Empirical.

IE 585 Introduction to Neural Networks. 2 Modeling Continuum Unarticulated Wisdom Articulated Qualitative Models Theoretic (First Principles) Models Empirical.

Similar presentations

Ads by Google