How to start with PVM

The contents

Setting up environment for PVM
Starting and stopping PVM
Setting up your home directory
Examples in PVM installation
How to compile and run PVM applications
Etiquette for users of PVM on workstations
Using PVM on IBM SP2 machine
How to use SP2
Troubleshooting
Hints on writing PVM code
Online books
xPVM - graphical interface to PVM

Setting up environment for PVM

If you use the csh (or tcsh) shell:

edit your .cshrc.local file to add the following to the end

setenv PVM_ROOT /s/pvm

setenv PVM_ARCH `$PVM_ROOT/lib/pvmgetarch`

setenv MY_PVM_BINARIES $HOME/private/pvm3/bin
set path=($path $MY_PVM_BINARIES/$PVM_ARCH $PVM_ROOT/bin/$PVM_ARCH $PVM_ROOT/lib $PVM_ROOT/lib/$PVM_ARCH)

Note that the apostrophes in the PVM_ARCH definition are back-ticks, not forward.

Log out and log in again to incorporate the new variables into your environment.

Back to the CS838-1 Home page Back to the contents

Starting and stopping PVM

Before you can run your PVM code, you must startup PVM daemons on the machines that you plan to use.
You can do that by typing
```
 pvm 
```
at prompt. This opens so called PVM console.
For example, if you are logged in to nova2 and want to run a PVM program on nova2+nova3, type
```
nova2$ pvm
```
If successful, you see a
```
 pvm> 
```
prompt. You're now running a one-node parallel virtual machine.

To create a two-node parallel virtual machine, type

pvm> add nova3
1 successful
		    HOST     DTID
                   nova3    c0000

Note that the machine you're on is automatically part of your PVM; it doesn't need to be explicitly added.
At this point you can simply run your PVM executable from the command line.
It's also possible to startup the local and remote PVM daemons through the PVM API, but for a small number of nodes it's probably easier to do it interactively.
Type
```
 help 
```
and the console gives you a list of commands for controlling your virtual parallel machine

Some useful commands to try:

conf	shows the current virtual machine configuration
add node	adds the node to your virtual machine
halt	halts (tears down) the virtual machine
quit	quits the virtual machine console, but doesn't tear down the virtual machine, so you can restart pvm later to manipulate your machine
spawn process	starts a program on your virtual machine

IMPORTANT: Stopping PVM

make sure that you stop PVM before you logout.
otherwise the daemons will keep running and could affect other users.
to stop PVM, use the PVM console
```
halt
```
command.
this will stop the daemons on all machines that you add-ed, and on the local machine.

nova1$ pvm
pvm> halt
nova1$

See also the Etiquette page for PVM users.

Back to the CS838-1 Home page Back to the contents

Setting up your home directory

Before you start to compile and run programs on your PVM configuration, I would recommend you to start with the following structure of your home directory, you can change it later by modifying your Makefile.

create directory
```
 $HOME/pvm3 
```
It will be used for your Makefile and PVM sources
in pvm3, create a soft link to PVM library by typing
```
ln -s /s/pvm-3.3.11/lib $HOME/pvm3/lib 
```
binaries will be stored in
```
 bin/$PVM_ARCH,
```
but this will be done automatically if you will use aimk on file Makefile.aimk

$HOME/------pvm3/------bin/-|-X86SOL2  - { exec files for Intel machines with Solaris}
        |          |        |-SUN4SOL2 - { exec files for Sparc machines with Solaris}
        |          |
        |          |--- lib/       { link to /s/pvm-3.3.11/lib/}
        |          |
        |          |--- Makefile.aimk
        |          |--- *.c

Back to the CS838-1 Home page Back to the contents

Examples in PVM installation

Whenever you do not understand some concept or PVM construction, consult the on-line PVM book .
The best way to learn about PVM programming is to start with examples in the installation.
In directory
```
/afs/cs.wisc.edu/s/pvm-3.3.11/src/pvm3.3.11/examples
```
you can find a couple of simple PVM source codes *.c
Read first the file Readme in this directory.
Make a detailed study namely of source codes of
```
master.c and slave.c. 
```
You will need to write a similar code, since:
typically, your program will consist of one master process whose task is to
- spawn slaves and
- assign them the initial portion of work (not necessarily input data, which can be generated or read from files locally by slaves),
- coordinate the activity of slaves if needed
  - in some algorithms, slaves behave totally independently upon master till the end (the queen problem for example)
  - in some others, the master periodically synchronizes slaves
  - in some others, the master is a central process which assigns work to slaves, receives resulting data and terminates the whole computation.

Back to the CS838-1 Home page Back to the contents

How to compile and run PVM applications

copy Readme and source codes *.c of examples into $HOME/pvm3
copy file Makefile.aimk into your $HOME/pvm3 directory. This is a modified Makefile from PVM installation, which worked for me on PVM examples
follow the instructions in Readme file to compile and execute the example codes
in the current setting, type for example
```
aimk hello
```
or any other example code name
modify the Makefile.aimk and play with compiling and executing other than example codes
If you plan to do multi-platform runs, use the PVM_ARCH environment variable in your Makefile.aimk to keep executables and object files for different architectures under separate directories.

Back to the CS838-1 Home page Back to the contents

Using PVM on IBM SP2 machine

Read the SP2 page for general information about using the SP2.

PVM works the same way on the SP2, but there's some complications that are due to our authentication setup in the department, and some limitations of kerberos 5 with multi-homed hosts.

If you don't need the communication speed of the switch, then using PVM on SP2 is simple:

login to one of the nodes (sp2-xx) using rlogin
startup and use PVM as you would on Solaris, using sp2-xx as host names

If you really need the speed of the switch, you will need to:

make your AFS home world-readable
```
 fs sa ~ system:anyuser rl 
```
create a directory in your AFS space that is also world-readable, and also world-writable if your PVM program will need to do some writing on the remote nodes. (shouldn't be needed)
edit/create
```
~/.rhosts 
```
and add entries for each sp2-nxx host. Use the fully qualified names (with the .cs.wisc.edu).
use the sp2-nxx names when starting up remote nodes

Consider changing permissions back to

 system:anyuser l

on those directories when you are not using PVM.

Back to the CS838-1 Home page Back to the contents

Troubleshooting

Some common problems, and possible fixes:

Remote node won't start
- Use
```
conf
```
  from the pvm console to make sure it's not already in your virtual machine
- Check if an old pvmd is running on the node; if so, kill it
- remove any /tmp/pvm* that belong to you on the remote node
- make sure your .cshrc.local script sets all the PVM* variables and path correctly
- if using the SP2 w/switch
  - make sure ~/.rhosts contains all the sp2-nxx host names
  - make sure your home directory and working directory are world-readable
- check /tmp/pvml.uid on the remote node for any interesting error messages

Back to the CS838-1 Home page Back to the contents

Hints on writing PVM code

design a parallel algorithm at the level of process graph in PVM.
debug, compile, and run on one machine,
if everything works, spawn the tasks on several machines or on a cluster of machines
measure the performance numbers and experiment with the scalability of the algorithm (running the program for various combinations of p and n)
tune the performance in case of detecting hot-spots
write a short report about the project with presenting the performance figures (speedups, efficiency, isoefficiency)

Back to the CS838-1 Home page Back to the contents

Online books and helps

Back to the CS838-1 Home page Back to the contents