Computer Science 400
Parallel Processing

Siena College
Fall 2008

Instructor:	Dr. James D. Teresco, Roger Bacon 338
Electronic mail:	domain: siena.edu, username: jteresco (best contact method)
Class URL:	[Link]
Class meetings:	Wednesday 3:35-6:15, Roger Bacon 302
Lab meetings:	Self-scheduled
Office hours:	Wednesday 1:30-3:00, by appointment

Course Information: [Complete Syllabus] [Lecture and Reading Schedule, Assignments]

Course News

We're done! Thanks for a great semester.
We have a small cluster! From the labs in Roger Bacon, its name is sccluster.teresco.org. From elsewhere, you'll probably have to connect first to a machine within the school of science network, then connect through to the cluster.
Your best way to use the cluster is from an X11 desktop, such as those found on the department's Linux systems. You will be able to open (and should definitely get in the habit of opening) multiple windows and things like emacs and your debuggers will display in their own windows. Here's the basic procedure:
- Open a terminal window on your Linux desktop
- Connect to the cluster with ssh and a special flag (-X) to set up X11 forwarding:
```
ssh -X sccluster.teresco.org
```
We have an excellent parallel debugger available on the Solaris nodes of our cluster.
- Since it is a graphical debugger, you'll need to log in using the -X flag as described above, then you'll need to include that -X when you connect to the appropriate Solaris node using ssh from the head node.
- For multithreaded programs, you invoke totalview by typing its name before your executable. For example:
```
totalview ./pthread_hello
```
- This will bring up a mostly self-explanatory interface that will allow you to determine where a program crashes, view different threads that are active in your program, set breakpoints, examine variable values, and a lot more.
- Note: In order for this or any debugger to be much help, you should compile and link your program with the -g flag, which will save symbolic information from the compile step in your executable, so the debugger can map the machine code and memory locations to source code lines and named variables in your program.
Running MPI jobs on the cluster
1. Compile and link your program with mpicc
2. Create "host file", a list of nodes on which to run your program, one hostname per line (a host file with all 5 of our FreeBSD nodes is available for your use in /cluster/nodes.freebsd).
3. Assuming the host file is called hosts and your executable is called mpiprog and you want to run 4 processes, you can use this command line:
```
mpirun -np 4 --hostfile hosts ./mpiprog
```
We have accounts in the RPI Computer Science Lab, giving us access to the Rensselaer Grid machines in Carlos Varela's Worldwide Computing Laboratory.
- One-time setup for MPI on your RPI CS account:
  1. Make sure you can log in to each host in the WCL with ssh and between hosts without having to issue a password.
  2. Create a file .mpd.conf in your home directory that contains the line
```
secretword=<secretword>
```
    where <secretword> is a word of your choice (not your password).
  3. Change the permissions on the file so only you have read access:
```
chmod 600 ~/.mpd.conf
```
  4. See if you can start an MPI daemon on your local host and run an MPI program:
```
mpd &
mpdtrace
mpdallexit
```
- When running your programs, you'll need to
  1. Create a file mpd.hosts that contains the list of hosts on which you wish to run your MPI program.
  2. Compile and link your MPI program with mpicc.
  3. Start the desired number of mpd processes with mpdboot -n <numprocs>.
  4. Run your program with mpdrun -np <numprocs> <yourprog> <args>.
  5. Alternately, use mpiexec to get more control over which processes are assigned to which nodes.
  6. When you are done running, kill your mpd's with mpdallexit.

Related Information and Links

If you are unfamiliar with Unix, please see the page of hints and Unix tutorial.
Very brief but very useful references, courtesy of Duane Bailey at Williams College:
- C
- Unix
- Emacs
RPI CS Lab Resources:
- Worldwide Computing Lab
- The Rensselaer Grid
Source code from our textbook's examples is available.

Computer Science 400 Parallel Processing

Siena College Fall 2008

Computer Science 400
Parallel Processing

Siena College
Fall 2008