PDRS: A Performance Data Representation System

1
PDRS A Performance Data Representation System

• Xian-He Sun
• Illinois Institute of Technology
• Sun_at_cs.iit.edu

Xingfu Wu (Northwestern)
2
Outline

• Introduction
• Design Framework
• Performance Database Module
• Relational Queries Module
• Summary

3
The SCALA (SCALability Analyzer) System

• Design Goals
• Predict performance
• Support program optimization
• Estimate the influence of hardware variations
• Uniqueness
• Designed to be integrated into advanced compiler
  systems
• Based on scalability analysis

4
The Structure of SCALA
5
What Is the Problem

• Evaluation relies on understanding of performance
  data
• The huge amount of data further complicates the
  issue
• Data manipulation functions depend on performance
  data organization and representation.
• Lack of appropriate models, representations and
  associated evaluation methods to understand data
  and locate bottlenecks

6
What Is PDRS

• A post-execution relational data system with
  novel performance data representation
• Perform knowledge diagnose
• Adopts the relational database, SQL, Java JDBC,
  visualization (Java3D) and auralization
  (JavaSound) technologies,
• Assumes the availability of modern compiler
  technology

7
Goal of PDRS

• Helps users identify the cause of the performance
  problem
• Assists programmers in optimizing the strategy or
  algorithm in their parallel programs

8
What Is New

• Relational data representation and management for
  advanced scalability, statistics, and symbolic
  analysis techniques
• Use a combination of performance visualization
  and auralization and object-oriented Java
  techniques

9
Design Framework of PDRS
10
Trace Data

• Collect performance data and store them with SDDF
• Use the SCALA instrumentation system (based on
  Vienna Fortran compilation system) to get the
  SDDF tracefile

11
Data Management Module

• Perform data filtering
• Preparation to store event history in a
  relational database form in PDRS
• Input is a SDDF tracefile
• Implemented in Oracle DBMS

12
Data management as a middle agent
13
Performance Database Module

• Classify performance data into five groups
• processor information
• memory information
• program information
• communication information
• I/O information
• Each class is an entity relation in the database
• An individual event is a tuple with a give
  identifier

14
Processor Information

15
Memory Information
16
Program Information
17
Communication Information
18
I/O Information
19
Relational Queries Module

• Retrieval via Relational Database Queries
• Symbolic Analysis
• Statistical Analysis
• Scalability Analysis
• Performance Model Generator

20

• The JDBC provides a bridge between Java
  applications (for this module) and performance
  database.

21
Symbolic Analysis

• Use symbolic evaluation to determine variable
  values, assumptions about and constraints between
  variable values, and conditions under which
  control flow reaches a program statement.
• Computations are represented as symbolic
  expressions defined over the program's problem
  and machine size.
• Each program variable is associated with a
  symbolic expression describing its value at a
  specific program point.

22
Statistical Analysis

• Determine code and machine effects, finds the
  correlation between program phases, identifies
  the scaling behavior of difficult-segments, and
  provides statistical performance data.
• Adaptive
• Bottleneck hypotheses
• Detailed analysis

23
Scalability Analysis

• Scalability combined with initial execution time
  can provide good performance prediction, in terms
  of execution times
• Crossing-point analysis finds fast/slow
  performance crossing points of parallel programs
  and machines
• Range comparison compares performance over a wide
  range of ensemble and problem size

24
Performance Model Generator

• Analysis/prediction based on different models.
• Crossing-point analysis model predicts execution
  time
• Empirical memory model infer CPU stall time due
  to memory hierarchy
• Probability model for non-deterministic variables
• Performance models given by the user

25
Performance Diagnostic Agent

• Provide performance advice identifying
  bottlenecks and comparison between measured and
  predicted performance
• Based on our approaches for symbolic,
  statistical, scalability analysis and performance
  models with certain diagnostic rules
• The function operation algorithm for this
  diagnostic module is as follows

26

• Performance analysis requests
• switch (analysis type)
• Statistical Retrieve the performance
  information required
• Get or compute the predicted performance
  range
• Compute the real result of requested
  performance parameter
• Compare the result with the performance range
• If (the result is not in the performance
  range)
• Give an explanation (using graphics and
  sound)
• break
• Scalability Retrieve the performance
  information required
• Get or compute the predicted scalability
  results
• Compute the real scalability results
• Compare the real result with the predicted
  results
• Explain the compared results (using graphics
  and sound)
• break
• Models Retrieve the performance information
  required
• Get the predicted performance range
• Compute the real result of requested
  performance parameter
• Compare the result with the performance range

27
Performance Visualization and Auralization

• Provide graphical display of performance
  information
• The basic goal is to use graphics and sound (Java
  2D, Java 3D and JavaSound) to display some advice
  and performance views about application programs.
  When performance bottlenecks occur, sound is used
  to inform users about some performance problem in
  their application programs

28
Summary

• Performance data representation using relational
  database
• Object-oriented programming
• Performance knowledge diagnosis
• Performance visualization and auralization
• A framework and on-going work
• Efficient filtering, mapping, database operation
• Implication rules