CrossGrid Performance Prediction Tutorial

1
CrossGrid Performance Prediction Tutorial
CrossGrid Tutorial Exercise PPC

Tomás F. Pena, José C. Cabaleiro, and Francisco
F. Rivera University of Santiago de Compostela,
Spain www.ac.usc.es/CrossGrid
2
Launching PPC GUI

• Open the PPC Viewer directly from Migrating
  Desktop
• Tools menu or PPC icon in main bar
• User data for the kernel to be predicted must be
  in the Virtual Directory in MD

3
PPC GUI main menu

• Submenus
• Grid get/set the Grid parameters
• Kernel select the kernel to be predicted
• Display show prediction results
• Help

4
First step Grid menu

• Getting/setting grid related information
• Three sources
• Data from actual grid monitoring (WP 3.3)
• Forecasted data (WP 3.3)
• Custom grid
• User can specify grid parameters such as number
  of nodes, latency and BW of links,

5
Get grid information from monitoring

• Get the information provided by JIMS
• Number of available nodes
• CPU models and IP addresses
• CPU loads

6
Show grid information

• Show the grid information
• Number of available nodes
• Bandwidth and latency between each pair of nodes
  data
• Information for each specific node
• CPU model and IP address
• CPU load

7
Step two Kernel menu

• Select the kernel to be predicted
• Kernels form WP 1, and other common ones
• Collective communications in MPICH-G2
• Iterative solvers for sparse matrices
• Matrix-vector multiplication
• In this exercise, iterative solvers is selected

8
Sparse matrix file

• For iterative solvers a square sparse matrix must
  be first loaded from a file
• The file must reside in the Virtual Directory in
  MD
• The matrix must be stored in Column Compressed
  Storage Format
• Other formats will be supported in future
  versions of PPC

9
Sparse matrix format

• The file must be in ASCII format with the next
  structure
• Number of columns (N) and number of non-zero
  entries (NNZ) (integers)
• A N1 elements vector with the position of the
  elements which start each column (integers)
• A NNZ elements vector with the row indexes of the
  non-zero entries (integers)
• Values of the non-zero entries, in a column wise
  fashion (optional)

10
Kernel menu Iterative Solvers

• Eight different iterative solvers
• This includes most of the standard stationary and
  non stationary iterative methods
• Solvers come from the Paraiso (Parallel Iterative
  Solvers) library (http//www.ac.usc.es/paraiso)

11
Selection of Iterative Solver parameters

• For each iterative method different scaling and
  preconditioner can be selected
• Some precontitioners need additional parameters,
  which can be introduced with the Parameters
  button

12
Step three Processors mesh selection

• Some kernels (iterative solvers, matrix-vector
  multiplication) are designed for a processors
  mesh
• The number of rows and columns of the processors
  mesh can be introduced with this menu

13
Step four Displaying results

• Different results of prediction can be displayed
• Number of computations for process (floating
  point operations)
• Number of communications for process (number of
  sends and receives)
• Predicted execution time
• Load balance
• Other kernel-specific displays

14
Floating Point operations

• The number of FLOPs per process
• The specific processor can be selected
• FLOPs for different parts of the kernel can be
  showed
• Iterative method itself
• Scaling of the matrix
• Computation of the preconditioner
• Application of the preconditioner
• Total FLOPs for the whole kernel

15
Predicted load balance

• Load balance is showed using a Kiviat diagram
• Load balance is expressed in function of
• FLOPs
• Necessary memory
• Computing time
• Three different Kiviat diagrams are displayed

16
For More Information

• Additional information in
• http//www.ac.usc.es/crossgrid/
• PPC is available from
• http//www.eu-crossgrid.org/