Running LAM on CS Lab machines

    Here's what to do:

     On some lab machine:

     % ssh-keygen -t dsa
       (accept the defaults, and just hit RETURN when asked for a passphrase)

     % cd ~/.ssh
     % cat id_dsa.pub  >> authorized_keys2    

     now try it out: ssh somemachinename 

    # if you use bash:  add to your .bashrc or .bash_profile file:
    export LAMRSH=ssh

    # if you use another shell (you likely don't) add to your .cshrc  file:
    setenv LAMRSH ssh

1. Boot lam
   To boot lam, you need to have a lamhosts file in your lam subdirectory, and run recon to see if lam can be started on all nodes in your hostfile:

    % recon -v lamhosts   # see the example below for this file's format 
        

   Then, to boot lam on the hosts in your host file (this has to be once at the beginning of each session you run mpi programs):

    % lamboot -v lamhosts 

2. Run MPI applications
   • Compile your application using mpicc for C code, and mpic+= for C++ code (mpif77 is the mpi compiler for Fortran):

      % mpicc -o myprog myprog.c 

   • Create an appfile describing which host will run which executable(s) In some cases you may have more than one executable in your mpi application (see the mandelbrot example below where there are separate master and slave executables). Use mpirun and the appfile to run your MPI application:

      
          % cat appfile
            # 18 a.outs across 6 nodes (will spawn 3 a.outs per node)
            n0-5 /home/newhall/lam/test/a.out 
            n0-5 /home/newhall/lam/test/a.out 
            n0-5 /home/newhall/lam/test/a.out 
          % mpirun -v appfile	 
          

     or use the -np # command line option to mpirun to specify the number of processes to start on the nodes. For more options, see the mpirun man page.

           % mpirun -np 25 /home/newhall/lam/test/a.out
          

   • After each individual run of an MPI application, it is good to run lamclean to clean up any residual state from the last run before starting a new run:

      % lamclean -v 

3. mpitask
   As your MPI application runs, you can look at the state of the tasks by running mpitask. To continuously run this program every X seconds, use watch:

          # every 1 second, run mpitask
          %  watch -n 1  mpitask
         

4. Shutdown lam
   After you are done running your MPI applications, you should shut down lam by running lamhalt, or if lamhalt doesn't completely clean things up, then try lamwipe:

    
        % lamhalt
        % lamwipe -v lamhosts 

  %  mkdir lam
  %  cd lam
  %  vi lamhosts	# add some hosts 
     
  %  recon -v lamhosts 
  %  lamboot -v lamhosts 
  %  tping -c1 N
  %  cp -r /usr/share/doc/lam3-dev/examples/main/mandelbrot .
  %  cd mandelbrot/
  %  gunzip *.gz
  %  mpicc -o master master.c 
  %  mpicc -o slave slave.c 
  %  vi appfile			# create appfile schema file: it describes
				# the application: each node and program
				# the node names come from 'recon -v lamhosts'
  %  cat appfile
     # 1 master, 5 slaves
     n0 master 
     n1-5 slave 

  %  mpirun -v appfile 		# run the MPI application
  %  xv mandel.out 		

  %  lamclean -v		# clean up lam stuff from this run
  %  mpirun -v appfile 		# re-run the mandelbrot app 
  %  lamhalt -v			# halt lam...you are done running lam apps
  %  lamwipe -v ../lamhosts     (if halt didn't work)