Description
For this and all future labs, you will be working on the CS Linux Lab system. If you are not physically in a Linux Lab (FAB 88-09 and 88-10), you may remotely log on to it: linuxlab.cs.pdx.edu. Download and unzip the file lab1.zip from D2L. You’ll see a lab1 directory with multiple versions of the sum program: sum.c sum-pthd.c sum-omp.c sum-mpi.c sum1.chpl sum2.chpl plus a couple of other files: Makefile, linuxhosts. 1 Set Up Your Environment • Check which shell you are using (bash, csh, ksh, or other): linux> echo $0 /bin/bash This info will be useful for setting up environment variables (see below). • Check that you have the latest version of gcc: linux> gcc –version gcc (Ubuntu 5.4.0-6ubuntu1˜16.04.4) 5.4.0 20160609 • Check that you have access to the MPI compiler, mpicc: linux> which mpicc /usr/bin/mpicc • Use addpkg to add the latest version of Chapel compiler to your environment: linux> addpkg … (a listing of available packages) Use arrow keys and the tab key to select chapel-1.16.0; then select and press enter. To effect the selection, you need to logout your session and re-login. 2 Compile and Run Pthreads Programs To compile Pthreads programs, use gcc with -pthread flag: linux> gcc -pthread -g -o sum-pthd sum-pthd.c The compiled programs are run just like regular C programs: linux> ./sum-pthd Note that for this sum-pthd.c program, the number of threads is hardwired in the program. In the future, we’ll see how to make that adjustable. 1 Exercise Add a printf statement in the worker() routine to show the id and the work range of each individual thread. You may want to place this statement under a control flag: #ifdef DEBUG printf(…); #endif This way, the same program can be compiled to two different versions: linux> gcc -pthread -g -o sum-pthd sum-pthd.c # regular version linux> gcc -DDEBUG -pthread -g -o sum-pthd-debug sum-pthd.c # debug version 3 Compile and Run OpenMP Programs To compile OpenMP programs, use gcc with “-fopenmp” flag: linux> gcc -fopenmp -g -o sum-omp sum-omp.c Again, the compiled programs are run just like regular C programs: linux> ./sum-omp Exercises 1. To see the non-intrusive nature of OpenMP, compile the program without the “-fopenmp” flag, and save it in a different target: linux> gcc -g -o sum-omp0 sum-omp.c The result is an identical copy to the sequential version. One way to confirm that this target code is different from the previous one is to generate and compare their assembly code: linux> gcc -fopenmp -S sum-omp.c # generate openmp code linux> gcc -S -o sum-omp0.s sum-omp.c # generate sequential code linux> wc sum-omp.s sum-omp0.s # compare their sizes linux> diff sum-omp.s sum-omp0.s # compare their contents 2. To further confirm that the program sum-omp is indeed running with multiple threads, insert a printf statement inside the for loop to print out the current thread id, which can be obtained by a call to omp get thread num(). For this to work, you also need to include the OpenMP header file: #include Question: How many threads are being used? 4 Compile and Run MPI Programs To compile MPI programs, use the command mpicc (which is a gcc wrapper): linux> mpicc -g -o sum-mpi sum-mpi.c Before running MPI programs, you need to setup a host file. Copy linuxhosts to your home directory, and set the following environment variable (different shells use different syntax): linux> export OMPI_MCA_orte_default_hostfile = ˜/linuxhosts # bash, ksh linux> setenv OMPI_MCA_orte_default_hostfile ˜/linuxhosts # csh, tsch 2 You should include this line in your shell startup file to avoid typing it in every time. For bash, ksh, csh, and tcsh, respectively, the file names are .bash profile, .kshrc, .cshrc, and .tcshrc. An MPI program is run with the command mpirun, with a flag -n <#copies> indicating the number of copies you’d like to execute: linux> mpirun -n 4 ./sum-mpi // running 4 copies of the program Note that for this program, the number of program copies is specified externally at the time of execution. Exercise Add a printf statement in sum-mpi.c to print out the values of two variables rank (current process id) and size (total number of processes). Compile and run the program to verify that four copies of the code are executed. 5 Compile and Run Chapel Programs Unlike the above three cases, where gcc handles all the compilations, to compile Chapel programs, a separate compiler, chpl, is needed: linux> chpl -g -o sum1 sum1.chpl linux> chpl -g -o sum2 sum2.chpl For running Chapel programs, you need to set the following env variables in your shell startup file: # for bash, ksh export GASNET_SPAWNFN=S # note: no space before and after = export GASNET_SSH_SERVERS=”bevatron boson …” # list of host names export SSH_CMD=ssh # for csh, tcsh setenv GASNET_SPAWNFN S setenv GASNET_SSH_SERVERS “bevatron boson …” # list of host names setenv SSH_CMD ssh The list of host names need to be manually copied from the file linuxhosts. A Chapel program is run with a flag -nl <#locales> indicating the number of locales (i.e. hosts) you’d like to use: linux> ./sum1 -nl 1 // running the program over 1 locale linux> ./sum2 -nl 4 // running the program over 4 locales Exercises 1. In both programs, the problem domain size N is a configurable constant. Try to change it at the time of execution: linux> ./sum1 –N=2000 -nl 1 2. Chapel view threads as a lower-level concept, hence does not provide a facility to show which thread a specific code piece in executed by. However, it does provide a facility to show locale information. Add a writeln statement inside the compute() function in sum2.chpl to show where the computation takes place. Use here.id to refer to the current locale’s name. Conclusion Summarize your experience with these four languages and tools. Which one is your favorite at this point? Remember your answer, and we’ll see if you’ll change your mind at the end of this course. 3