Performance evaluation using GPM

1
Performance evaluation using G-PM
CrossGrid WP 2.4.1 Tutorial

Wlodzimierz Funika, Tomasz Arodz, Marcin
Kurdziel ACC Cyfronet AGH, Kraków, Poland
2
Interpretation of Performance Data

• In principle, we can do interpretation by
  comparing the values presented in displays across
  sites, nodes, processes, fragments of code, time
  intervals, with data on infrastructure in
  background
• You can use data summarized over time vs.
  current values, sometimes the former suffices to
  discover a bottleneck in the program, when it is
  not enough we explore it in more details
• For more advanced performance evaluation we use
  user-defined metrics, especially when dealing
  with interactive applications
• But prior to all these actions you should plan an
  idea what you want to measure and how to attack
  the data provided by the measurements
• Therefore you should bear in mind that you have
  to possess a coarse- or fine-grained knowledge of
  the applications operation.

3
Introduction to the G-PM tool

• The G-PM tool is provided for graphical
  performance analysis of parallel grid
  applications
• Support for measurement of
• MPI-related quantities
• Other application-related quantities
• Grid infrastructure quantities
• A stand-alone X-Windows tool
• Based on the OCM-G monitoring system

4
Outline of the perf. evaluation actions

• Starting the OCM-G and the application
• Starting the G-PM
• gt./gpm ApplicationID
• Realizing Performance Evaluations
• Definition of the measurement
• Definition of the visualization
• Observation of the performance data
• For New , Change, Delete functions, etc. please
  goto p. 1 or 2, correspondingly

5
Defining a measurement

• Choosing a type of measured quantity (a metric)
• e.g. data volume sent with MPI_Send()
• Choosing the processes/hosts of interest
• e.g. source of the MPI communication (i.e. that
  from the parameters of MPI_Send ())

• Choosing the partner processes/hosts of interest
• e.g. destination of the
• MPI communication
• (i.e. that from within
• the parameters of
• MPI_Send ())

6
Defining the visualization

• Choosing a visualization type
• e.g. MultiCurve diagram
• Choosing the parameters of scale
• Linear scale
• Variable scale range
• Setting Update interval at which the Measurement
  is realized
• Adjusting the measurement accuracy to an
  acceptable overhead

7
Observing the performance data

• Each curve represents a measurement
• Send volume of communication between a selected
  source and destination process within a time
  interval

8
Knowledge learnt

• How to submit the application so that its
  performance can be monitored by G-PM
• How to start the G-PM tool for a given
  application
• How to move throughout the major steps in
  preparing the measurement and visualization

9
For More Information

• Introduction to the G-PM main concepts
• M. Bubak, W. Funika, R. Wismüller.
• The CrossGrid Performance Analysis Tool for
  Interactive Grid Applications.
• In Recent Advances in Parallel Virtual Machine
  and Message Passing Interface.9th European
  PVM/MPI Users Group Meeting.
• Linz, Austria, Sept. 2002, LNCS 2473,
  Springer-Verlag.
• Detailed G-PM installation instructions
• http//www.eu-crossgrid.org/Deliverables/1stYear-
  revised_delivarables/
• CG2.4.1-D2.3-v1.0-CYF001-PrototypeDoc.pdf
• The G-PM tool is available from
• http//www.eu-crossgrid.org/

10

• Thank you !