Presentation is loading. Please wait.

Presentation is loading. Please wait.

Performance Analysis And Visualization By:Mehdi Semsarzadeh Chapter 15.

Similar presentations


Presentation on theme: "Performance Analysis And Visualization By:Mehdi Semsarzadeh Chapter 15."— Presentation transcript:

1

2 Performance Analysis And Visualization By:Mehdi Semsarzadeh Chapter 15

3 Outline : Introduction Grid Performance Problems Grid Application Analysis Constraint Grid Performance Analysis Techniques Current Performance Analysis Tools

4 Introduction Performance measurement techniques provide the raw data needed to identify and correct performance problems Performance visualization, correlation, evaluation tools must process raw performance data, correlate it with appropriate network and highlight performance problems in meaningful ways

5 Grid Performance Problems performance measurement at many levels

6 Grid Performance Problems The raw data must be correlated across semantic levels and present in way that match the user’s semantic model interactive drilldown to lower semantic levels, allowing users to explore underlying causes of poor performance Dynamic optimization during application execution

7 Grid Application Analysis Constraint Complex, High-Dimensional Data Performance analysts must capture dynamic performance data at all system levels The components of these levels interact on a wide variety of time scale (from ns for HW to second or minutes for interacting with network devices) The number of components can be very large (100 or 1000 processors) =>Performance analysis tools must rely on strategies that highlight key relationships rather than on multivariant displays that show everything while illuminate nothing

8 Grid Application Analysis Constraint Multilevel Semantic Correlation The wider semantic gap between user-written code and HW /SW resource, by increase in application sophistication, (e.g. loop transformation by compiler) Different execution behavior (from what developer expected) because of aggressive compiler transformations => supporting hierarchical HW/SW representation to identify the true root cause

9 Grid Application Analysis Constraint Mixed Ordinal and Categorical Data –Ordinal Data: the number of messages sent or received, the number of cache misses, … –Categorical Data: processor states (e.g., blocked, queued, idle, or busy), software state, …. The categorical data lackes a total ordering It cannot be directly mapped to the same visual attributes as ordinal data => using symbols/colors to identify individual categorical values =>mapping categorical values to numerical values

10 Grid Performance Analysis Techniques Dynamics Static Dynamic Cardinality Univariate Multivariate Ordinality Ordinal Semiordinal Categorical

11 Grid Performance Analysis Techniques Multivariate Data –scatter plots –kiviat Diagrams

12

13 Eight Processor Utilization With Kiviat Diagram

14 Grid Performance Analysis Techniques Semiordinal Data –Histograms use to provide relative counts of operation Network traffic x-axis:domain names y-axis:message retransmition count –GanttCharts usually use to display processor utilization

15 Grid Performance Analysis Techniques

16 Structural Display –correlates dynamic performance data with structural representation of HW/SW component –Geographic and network mappings –Architectural mappings –Execution graphs –Source code correlation

17 Geographic and network mappings

18 Execution graph

19 Source code correlation

20 Current Performance Analysis Tools ParaGraph IPS-2 and Paradyn Chitra Pablo and Svpablo Avatar

21 Paradyn Relies on a hierarchical approach The hierarchy is represented as a tree The root is the program and branches represent machines and processes Bottleneck identification by supporting real-time insertion and removal of measurement probes during execution. Use w3 technique to automatically determine the causes of bottleneck –Why is the application performance poorly? –Where is the performance problem? –When does the problem occur?

22 Paradyn

23

24

25 Avatar A virtual reality framework for analyzing complex, time- varying, multivariate data Displays a three-dimensional globe, the analyst can rotate the globe, view traffic on a worldwide basis The analyst can select a node to obtain detailed information

26 Avatar

27 Grid Performance Analysis Challenges Application codes become more complex,dynamic =>Their behavior is increasingly nondeterministic the posterior analysis becomes ineffective => performance optimization softwares must support realtime correction of performance problems =>closed-loop performance optimization: distributed sensors that provide low-overhaed,are accessible from any site on the grid resource actuators that enable tasks to modify resource allocation policies a set of fuzzy logic decision mechanism to realize system modification

28 References The Paradyn Parallel Performance Measurement Tools –Barton P. Miller Mark, D. Callaghan, Jonathan M. Cargille, Jeffrey,… Real-Time Geographic Visualization of World Wide Web Traffic –Stephen E. Lamm and Daniel A. Reed www-pablo.cs.uiuc.edu


Download ppt "Performance Analysis And Visualization By:Mehdi Semsarzadeh Chapter 15."

Similar presentations


Ads by Google