Running HPF Programs on the NOW

The following is a description of how to run HPF programs on the Berkeley NOW.

Setting up your environment

Type:

source /usr/castle/proj/cs267/cs267.cshrc

Log into the NOW

Log into the NOW as described in assignment2

Compiling a program

You'll now be able to compile a sample program. Copy the following file to a directory: karp.hpf

This is a hpf file that computes PI through integration.

Now, execute the following command to compile the program. By default, the compiler will add the libraries necessary to communicate via sockets. There are many other compiler options listed in the pghpf's web pages.

pghpf -Mstats karp.hpf -o karp

This will compile the karp program, and insert profiling code that will assist in reporting the memory used/messages sent by the program. The compiler will produce a file called 'karp'

If you experienced problems trying to compile, make sure that your path was set up correctly as noted in the first section above.

Running a program

Setting things up

There are two ways to execute programs - securely and unsecurely. We strongly recommend that you use the secure method.
Securely

Set up kerberos
rkinit to the remote machine that you are running on: (You don't have to rkinit to all remote machines, since the rsh command will transfer the ticket to the other remote machines.)

This will set up a ticket on the current machine you are running on. You should only run 'kinit' to create tickets on the machine that you are typing at, otherwise your kerberos password will be transfered over the network in the clear. At the machine you are typing at, you should run 'rkinit' to install tickets on machines remotely - rkinit will not issue your password in cleartext.

For example, if you are typing at the machine whenever.cs.berkeley.edu and you wanted to run programs from the machine u0.cs.berkeley.edu, you would type:

rkinit u0.cs.berkeley.edu

Unsecurely (not recommended)

Add the current hostname to your .rhosts file

For example, if you are logged into u0.cs.berkeley.edu, then add

u0.cs.berkeley.edu

to the end of your ~/.rhosts file.

Running the program

On one processor:

Type ./karp

Run it on 4 processors

SECURELY: ./karp -pghpf -np 4 -host remote_host1,remote_host2,remote_host3 -stat alls
UNSECURELY: ./karp -pghpf -np 4 -host remote_host1,remote_host2,remote_host3 -stat alls -rsh /bin/rsh

where remote_host1,remote_host2,remote_host3 are the names of 3 machines that you are going to run on, e.g. u1,u2,u3

Note that HPF will automatically run a copy on the current host, so you specify only the N-1 remote hosts on the command line.

**TODO: create a 'rsh'-like shell to spawn jobs using GLUnix. HPF expects a rsh tool to spawn off jobs.

Runtime options

Consult the HPF documentation for the entire list of options, but here are some common ones:

-v for verbose runtime
-dyn for dynmically pick nodes based on load average

Advanced

The HPF compiler also supports some of CM FORTRAN. By default, CM FORTRAN is accepted in some cases. See the HPF documentation for more details.

HPF on MPI

Compiling a HPF program with MPI

First, setup your environment as to compile a socket version of HPF program.

To compile with MPI option, you have to setup two environment variables before compiling:

setenv HPF_MPI "/disks/barad-dur/now/MPI/mpich/lib/solaris/hpf/"

setenv HPF_SOCKET "-lglunix -lLAM -lsocket -lnsl -lposix4 -lLanaiDevice -lbfd -liberity -lthread -lm"

Then, compile your program with the -Mmpi option:

pghpf -Dmpi your_program.hpf -o your_executable

Running the executable

Setup the environment for the MPI. For more information consult the MPI on NOW HowTo

Then, run your executable as follows:

your_executable -pghpf -np #_of_nodes

Note: MPI command line arguments that defined in MPI on NOW HowTo can still be used.

chad@cs.berkeley.edu