Performance Models for Parallel Applications in Grids

1
Performance Models for Parallel Applications in
Grids
Presented By Sanjay.H.A Ph.D Student

• Research Supervisor
• Dr. Sathish Vadhiyar

Grid Application Research Lab Super Computer
Education and Research Center Indian Institute of
Science
2
Outline

• Introduction
• Rational Polynomial Model (ICPP 2006)
• Motivation and Objective
• Proposed Model
• Experiments and Results
• Limitations
• Adaptive Performance Model
• Motivation
• Automation Procedure
• Experiments and Results
• Procedure for reducing the Modeling Overheads
• Conclusions and Future Work

3
Introduction

• Performance Model
• Predicts Execution time of an application
• provide insight into the performance
  relationships between an application and the
  system used for execution
• Serves as objective function for system Resource
  Negotiator/Scheduler
• Identification of Bottlenecks in Applications and
  System
• Fine tuning of Algorithms
• Approaches
• Trace-event Simulation
• Parameterized analytical models
• Curve-fitting models

4
Outline

• Introduction
• Rational Polynomial Model (ICPP 2006)
• Motivation and Objective
• Proposed Model
• Experiments and Results
• Limitations
• Adaptive Performance Model
• Motivation
• Automation Procedure
• Experiments and Results
• Procedure for reducing the Modeling Overheads
• Conclusions and Future Work

5
Rational Polynomial Model

• Presented at ICPP 2006 held at Columbus, Ohio,
  USA.
• Motivation
• Grid / Distributed Environment
• Provides resource required to execute a
  scientific application
• Shared by multiple users
• Performance of application may be impacted in
  dynamic and often unpredictable ways
• Scheduler needs Performance model to match the
  applications and machines
• Curve-fitting modeling strategies needs lesser
  information about application
• Existing Curve-fitting Modeling Strategies assume
  uniform Loading Conditions
• This Assumption is Unrealistic in
  Non-Dedicated Environments

6
Objective

• To Analyze and Build a Curve-fitting model that
  attempts to predict execution times for any load
  conditions that may exist on the systems during
  application execution.

7
Rational Polynomial Model (cont ..)

• We evaluate the model with the help of ScalaPACK
  eigen value problem
• Measurement and prediction tool
• Network Weather Service (NWS)
• Loading Parameters
• Available CPU
• Available Bandwidth
• Models were evaluated and compared based on
• Average percentage prediction errors

• Trend shown by the actual and predicted values

8
Proposed Model

• General Model
• Computation and Communication Complexities
• Predicting execution times of generic parallel
  application
• Non-dedicated environments with large load
  dynamics
• Measure available CPUs and bandwidth at periodic
  intervals from the beginning to end of the
  application execution
• Take average of CPUs and bandwidth collected for
  each machine
• Avg_avail_cpu avg_avail_band
• Take minimum of all the avg_avail_cpu and
  avg_avail_band
• Min_avg_avail_cpu min_avg_avail_band

9
Proposed Model(cont..)

• Training
• Problem size, min_avg_avail_cpu ,
  min_avg_avail_band and Execution time
• Prediction
• Min_avg_avail_cpu and min_avg_avail_band cannot
  be measured before the application execution
• Forecast the min_avg_avail_cpu and
  min_avg_avail_band values based on the history
• NWS forecasting tool

10
Proposed Model(cont..)

• Coefficients
• Polynomial fit using data points in the training
  range
• Linear regression problem
• Axb
• Prediction of execution time
• Coefficients with system and application
  parameters are used in the model.

11
Experiment and Results

• Applications
• ScaLAPACK parallel eigen value problem
• Parallel Conjugate Gradient (CG) from NPB
• Parallel FFT application from FFTW package
• System Specifications
• 8-processor Intel Pentium IV
• 32-processor IBM P720
• Monitoring
• Available CPUs of the nodes and available
  bandwidths of inter node links are collected
  every 2 minutes
• Load Dynamics
• Random CPU and Network loading during application
  execution

12
Load Dynamics
Network Load
CPU Load
13
ScaLAPACK Eigen Value

• Intel System
• Computation complexity cubic
• Communication complexity quadratic
• Training 3000 7000
• Prediction 7500-12000
• Compared our model with 3 different models

14
Prediction for Eigen value problem on 8 Intel
Processors
Our Model 11.86 Prophesy
16.42 Type1 Multi-variate 20.15 Type 2
Multi-variate 13.84
Avg. Percentage Prediction Error on IBM System
15
15
Limitations

• Model does not considers number of processors as
  input
• User has to provide approximate complexities for
  the model
• Data required to train the model is large
• Performance model considered here is static

16
Outline

• Introduction
• Rational Polynomial Model (ICPP 2006)
• Motivation and Objective
• Proposed Model
• Experiments and Results
• Limitations
• Adaptive Performance Model
• Motivation
• Automation Procedure
• Experiments and Results
• Procedure for reducing the Modeling Overheads
• Conclusions and Future Work

17
Adaptive Performance Model

• Motivation
• User is unaware of characteristics of parallel
  application , which he/she submits to grid.
• User submits the application with different
  problem size and different number of processor at
  different time
• Scheduler has to take dynamic scheduling decision
  efficiently.
• Performance model has to take care of change of
  system parameters dynamically .
• Automatically determines approximate complexities
  and builds
• Adaptive Performance Model

18
Automation Procedure
Prediction and Model Evaluation phase
19
Automation Procedure Cont..

• Metric for Goodness of fit

• Residual Sum of Square
• Error Variance
• Standard Error

• Training Experiments
• Design training range in terms of problem size
• Conduct training experiments on single processor
  for single processor training range
• Conduct training experiments on 2,4,and 8
  processors for a designed training range

20
Automation Procedure Cont..

• Computation Modeling in terms of problem size
• Input Predefined function, CPU functions and
  single processor training data.
• Evaluate the all predefined functions and CPU
  combination with training data
• Top Computation functions are chosen considering
  the metric standard error.
• Output List of top Computation Complexity in
  terms of problem
• size and CPU
• Communication Modeling in terms of problem size
• Input Predefined functions, Bandwidth
  functions, Computation
  complexity list, 2 processor training data
• Fixing Computation complexity, evaluate
  predefined communication functions and bandwidth
  combinations
• Top computation and communication models are
  chosen considering the metric standard error.
• Output List of top Computation and
  Communication Complexities with
  problem size, CPU and bandwidth

21
Automation Procedure Cont..

• Tuning the Model in terms of number of processor
• Input Processor functions, 2,4 and 8 processor
  training data, Top Computation and
  Communication Models
• Evaluating the computation and communication
  Models with processor functions.
• Top computation model with processor functions
  are chosen Considering the metric standard error.
• Fixing the Computation model with problem size,
  CPU and processor , top communication models with
  processor functions are chosen
• Output List of top Computation and
  Communication Complexities with problem
  size,processor, CPU and bandwidth

22
Automation Procedure Cont..

• Prediction and Model evaluation Phase
• User Input problem size, Number of processor
• Min_avg_avail_cpu and min_avg_avail_bandwidth are
  predicted using NWS forecaster by feeding past
  cpu and bandwidth values
• Execution time is predicted using top ranked
  model.
• After execution of that task, data will be added
  to training set
• To address the dynamic change of system
  parameters we evaluate the model list with
  training data and change the model rankings
• For every function evaluation we are deleting
  least ranked functions using percentile
  technique.
• Automatically removing the training data which
  contains anomalies.

23
Experiments and Results

• Applications
• ScaLAPACK parallel eigen value problem
• Parallel Conjugate Gradient (CG)
• Parallel FFT application from FFTW package
• Integer sort
• Molecular Dynamics
• Poisson Equation in 2-D jacobi Decomposition
• All pair shortest path
• 8-processor Intel Pentium IV
• 200 random problem size and processor pair are
  chosen for prediction

24
Experiments and Results Cont..
Molecular Dynamics 15
Integer Sort 11
25
Procedure for reducing the Modeling Overheads

• Previous technique 3 hrs (18000 iterations)
• Procedure
• Predefined functions 4 groups
• List of top functions
• 3 Representative Models for Last three groups
• Template final list for 3 representative models
• Current Technique 56 Min (1/3 of the previous
  technique)
• Results
• Parallel FFT 20
• Parallel Integer Sort 11
• Computation Phase
• Initially Polynomial and logarithmic functions
  are evaluated
• Prepare the list of top functions whose Std. Err.
  is within the threshold limit
• Remaining predefined functions are splitted into
  3 groups that belongs to 3 polynomial models.
• If the representative model is in the list then
  that function group will be evaluated

26
Procedure for reducing the Modeling Overheads

• Communication Phase
• For the computation model in the list , evaluate
  all polynomial and logarithmic communication
  functions with bandwidth functions,
• Top function list
• Template bandwidth functions list will be ready
  for 3 representative models
• If the representative model is in the list then
  that group will be evaluated with template
  bandwidth function
• Template final list will be ready if computation
  model belongs to any of 3 polynomial models
• Next computation model will be evaluated. If its
  template final list is ready, directly template
  final list will be evaluated.
• Model in terms of Processors
• For each comp_comm_model , evaluate with
  processor functions
• If the computation and communication models
  belongs to representative models then prepare the
  template list for that combination
• For the next comp_comm_model if the template list
  is ready, then only those functions will be
  evaluated.

27
Outline

• Introduction
• Rational Polynomial Model (ICPP 2006)
• Motivation and Objective
• Proposed Model
• Experiments and Results
• Limitations
• Adaptive Performance Model
• Motivation
• Automation Procedure
• Experiments and Results
• Procedure for reducing the Modeling Overheads
• Conclusions and Future Work

28
Conclusion

• We developed the automatic prediction technique
  that
• Determines approximate complexities for the model
• Adapts Performance Model for any loading
  condition
• Satisfies accuracy.
• We evaluated our technique with 7 parallel
  application
• We also proposed the technique to reduce the
  overhead of automation technique.
• In all cases the model gave good predictions of
  execution times

29
Future Work

• To Develop systematic methods to determine the
  approximate complexities of any multiple
  parameter application automatically
• To develop a smart scheduler which takes
  efficient decision based on our performance model
• To Augment our techniques for predicting
  execution time for complex multi-phase and
  multi-component applications
• To Extend our work to include I/O related
  parameters for predicting the behavior of I/O
  intensive scientific application.

30
References

• Jay Yagnik, H.A Sanjay, Sathish Vadhiyar.
  Performance modeling based on multidimensional
  surface learning for performance prediction of
  parallel application in Non-Dedicated
  enviornament, ICPP-2006,pages-513-520,
  Columbus,Ohio,USA
• V. Taylor, X. Wu, J. Geisler, X. Li, Z. Lan, M.
  Hereld, I. Judson, and R. Stevens. Prophesy
  Automating the Modeling Process. In Proceedings
  of the Third Annual International Workshop on
  Active Middleware Services, pages 3-11, Tokyo,
  Japan, August 2001.
• V. Taylor, X. Wu, and R. Stevens. Prophesy An
  Infrastructure for Performance Analysis and
  Modeling of Parallel and Grid Applications. ACM
• SIGMETRICS Performance Evaluation Review,
  30(4)13-18, March 2003.
• J. Schopf. Structural Prediction Models for
  High-Performance Distributed Applications. In
  Proceedings of the Cluster Computing Conference
• (CCC '97), Atlanta, U.S.A., March 1997.
• J. Schopf and F. Berman. Using Stochastic
  Information to Predict Application Behavior on
  Contended Resources. International Journal on
  Foundation in Computer Science, 12(3)341-364,
  June 2001.
• P. Dinda. Online Prediction of the Running Time
  of Tasks. In Proceedings of the 10th IEEE
  International Symposium on High Performance
  Distributed
• Computing (HPDC-10'01), pages 383-394, San
  Francisco, U.S.A., August 2001.
• DataFit. http//www.curvefitting.com/datafit.htm

31
THANKYOU !
Questions ???