Next: Architecture selection field Up:
Help Previous: Mesh / Hypercube
Performance modelling
What it is
The purpose of this applet is to demonstrate influence of system parameters
on the performance of the divisible load processing.
By changing the ratio of startup time to processing rate (S/A) amd communication rate to processing rate (C/A)
one can instantly observe the changes in the shape of processing time T(V), speedup S(m), and efficiency E(m) curves.
Usage
The applet contains six parts:
- architecture selection field
- S/A and C/A ratio selection field
- current values of S/A, C/A
- Function of processing time T in problem size V
- Function of efficiency E in processor number m
- Function of speedup S in processor number m
Quick help
You must select some architecture first, by clicking on its name,
to see the performance functions in parts (4), (5), (6).
Drag the marker in field 2 with your mouse, to change the ratios S/A, C/A.
The scale can be changed from logartihmic to linear in panel 4 by double-clicinkg on it
More info
There are six architectures to chose from in part (1) considered:
- chain (with or without communication coprocessors)
- star (with or without communication coprocessors)
- 2-D mesh (only with communication coprocessors) with p=4 (number of
simultaneously used communication ports) and Peters - Syska data distribution
algorithm
- hypercube (only with communication coprocessors)
It is assumed for every architecture that processors and communication links
are identical and described by A,C and S parameters.
Input data for divisible task model comprises:
- A - inverse of processing speed (time needed to process one data
unit)
- C - inverse of communication speed (time needed to send one data
unit)
- S - startup time (communication delay incurred during the
initialization of the connection)
- V - size of problem (number of data units)
- m - number of processors (only for chain and star)
- p - number of simultaneously used communication ports (only for
mesh)
- k - number of moves in Peter - Syska algorithm = number of layers
- 1 (only for mesh)
- d - dimension - number of communication links for one processor (only
for hypercube)
This applet allows for changing some of those values.
The rest of them are already set.
- for chain, star and hypercube: A=1, m=2,4,8,16,32,64,
- for mesh: A=1, m=1,5,25,125,625, p=4.
Problem size V is between 1 and 1E+12 for all architectures.