CS 584 Lecture 17

1
CS 584 Lecture 17

• Assignment?
• C program
• Papers
• Test
• When?

2
Glenda Help

• Glenda tutorial
• Postscript paper
• Local configuration
• HTML document
• Both available from our web page

3
Glenda

• Supports 2 models of parallelism
• Agenda
• Master-slave
• SPMD kindof
• Master just spawns jobs and waits for them to
  finish

4
Glenda goals

• Be Linda
• Hide PVM as much as possible
• Maintain PVM message passing ability
• Maintain PVMs portability

5
Linda vs. Glenda

• Linda has no mechanism to determine a process
  number.
• Linda provides no means to communicate other than
  tuple space.
• Glenda does not provide the eval function
• We must spawn the process, and put out a tuple
  for it to evaluate.

6
Glenda functions

• tid gl_mytid()
• tid gl_spawn( char name )
• gl_out( char key, ..)
• gl_in(char key, ..) similarly gl_inp
• gl_rd(char key, ..) similarly gl_rdp
• gl_outto(int tid, char key, ..)
• gl_into(char key, ..)
• gl_exit()

7
Glenda programming

• First join Glenda by calling gl_mytid
• Then the master uses gl_spawn to start up as many
  workers as you need.
• One per call
• The tid of the worker is returned
• gl_outto and gl_into are used to send and receive
  tuples directly
• No Structures, unions, or typedefs

8
gl_out

• Variable argument list
• Values can be scalar or arrays
• Array sizes are implicit unless declared
• Sizes of 2-d arrays must be declared

gl_out( "data", j, k, val) gl_out( "row", l,
xlen) // declares the length to be
len gl_out( "col", xjlen) // 2d array
9
gl_in, gl_inp, gl_rd, and gl_rdp

• Similar to gl_out
• gl_in blocks gl_inp returns 1 or 0
• The ? is used to indicate variable data
• Other arguments used for matching

gl_in( "data", j, k, ? val) gl_in( "row", l, ?
xlen) // declares the length to be
len gl_in( "col", ? xj) // 2d array
or single element
10
gl_outto and gl_into

• Send a tuple directly to a process
• Similar to gl_out and gl_in
• gl_out requires the destination id

gl_outto( tid, "data", j, k, val) gl_into(
"data", j, k, ? val)

• If tid is an array, gl_outto is a broadcast.

gl_outto( tid len, "data", j, k, val)
11
How does Glenda work?

• Runs on top of PVM
• Special process
• Global Tuple Server
• Glenda uses a preprocessor to convert a Glenda
  program into a PVM program

12
Using Glenda

• Add snell/glenda/bin to your PATH
• Copy the global tuple server to your
  pvm3/bin/HPPA directory.
• Write your Glenda program
• Must have a .cg extension
• Run cgpp on your program
• Compile the result using the C compiler

13
Using Glenda

• Copy your executables to your pvm3/bin/HPPA
  directory
• Start up PVM
• Configure your virtual machine
• Run gts
• Run your master program

14
Assignment

• Redo Lab 1 using Glenda
• Compare your speedups and execution times with
  the previous two labs.