MPICH Test Suite This test suite is a *supplement* to other test suites, including the original MPICH testsuite, the Intel testsuite, and the IBM MPI test suite (or test suites derived from that test, including the MPI C++ tests). Building the Test Suite ======================= In many cases, configure will find the MPI implementation automatically. In some cases, it will need some help. For example: For IBM MPI, where the compilation commands are not mpicc and mpif77 etc.: ./configure CC=xlc MPICC=mpcc F77=xlf MPIF77=mpxlf CXX=xlC \ MPICXX="mpCC -cpp" FC=xlf90 MPIFC=mpxlf90 \ --disable-spawn \ --enable-strictmpi (or the _r versions of the compilers) If mpicc and friends are not in your default path (and you do not want to add them), you can specify the path with --with-mpi=. For example, if they are in /usr/local/mympi/bin, use ./configure --with-mpi=/usr/local/mympi (configure will append the bin to the path that you give). You may need to add MPI_SIZEOF_OFFSET=8 . The option "-cpp" is needed for at least some versions of mpCC to define the C++ bindings of the MPI routines. For implementations that do not implement all of MPI-2, there are --disable options, including --disable-spawn and --disable-cxx. To restrict tests to just what is defined in the MPI specification, use --enable-strictmpi . The script that runs the tests assumes that the MPI implementation supports mpiexec; you should consider this the first test of the implementation. Setting Options =============== The following environment variables will modify the behavior of the tests MPITEST_DEBUG - if set, output information for debugging the test suite MPITEST_VERBOSE - if set to an integer value, output messages whose level is at least that value (0 is a good choice here) MPITEST_RETURN_WITH_CODE - Set the return code from the test programs based on success or failure, with a zero for success and one for failure (value must be yes, YES, true, or TRUE to turn this on) MPITEST_THREADLEVEL_DEFAULT - Set the default thread level. Values are multiple, serialized, funneled, and single. Batch Systems ============= For systems that run applications through a batch system, the option "-batch" to the runtests script will create a script file that can be edited and submitted to the batch system. The script checktests can be run to summarize the results. Specifically, (assuming the bash shell, and that the directory "btest", a subdirectory of the test suite directory, is used for running the tests): export MPITEST_BATCHDIR=`pwd`/btest runtests -batch -tests=testlist ... edit btest/runtests.batch to make it a value batch submissions script ... run that script and wait for the batch job to complete cd btest && ../checktests If a program other than mpiexec is used in the batch form to run programs, then specify that to runtests: runtests -batch -mpiexec=aprun -tests=testlist (Here, aprun is the command used on Cray XE6 systems.) Note that some programs that are used to run MPI programs add extra output, which can confuse any tool that depends on clean output in STDOUT. Since such unfortunate behavior is common, the option -ignorebogus can be given to checktests: cd btest && ../checktests --ignorebogus See "More control over running tests" to see how to control how many processes per node on used. For example, on a Cray XE-6, this command line to runtests can be used: runtests -batch -tests=testlist -ppnarg="-N %d" -ppn=2 -showprogress \ -mpiexec=aprun This runs at most 2 processes per node. Note that this can take a long time to execute because it builds all of the executables required for the tests (over 800 of them!). The "-showprogress" flag lets you know that something is happening, but is not necessary. Controlling the Tests that are Run ================================== The tests are actually built and run by the script "runtests". This script can be given a file that contains a list of the tests to run. This file has two primary types of entries: directories: Enter directory and look for the file "testlist". Recursively run the contents of that file program names: Build and run that program Lines may also be commented out with "#". The simplest program line contains the name of the program and the number of MPI processes to use. For example, the following will build the program sendrecv1 and run it with 4 processes: sendrecv1 4 In addition, the program line can contain key=value pairs that provide special information about running the test. For example, sendflood 8 timeLimit=600 says to build and run the program sendflood with 8 MPI processes and permit the test to run for 600 seconds (by default, at least for MPICH, the default timelimit is 180 seconds). Other key=value pairs can be used to select whether a program should be run at all, depending on the abilities of the MPI implementation (this is particularly important for Fortran programs, since preprocessor support for Fortran is a non-standard extension to the Fortran language, and there are some compilers that would not accept Fortran programs that used the preprocessor). The most important key=value pairs are: timeLimit=n : Use a timelimit of n seconds arg=string : Run the program with string as an argument to the program mpiexecarg=string : Run the program with string as an argument to mpiexec env=name=value : Run the program with environment variable "name" given the value "value" mpiversion=x.y : Build and run the program only if the MPI version is at least x.y. For example, distgraph1 4 mpiversion=2.2 will build and run distgraph1 with 4 MPI processes only if the MPI version is at least 2.2. strict=bool : If bool is false, only build and run the program if --enable-strictmpi was not used in configuring the test suite. That is, a line such as neighb_coll 4 strict=false Says that this test is not valid for a strict MPI implementation; it contains extensions to the standard, or in the case of some MPICH development, MPIX routines resultTest=proc : This is used to change the way in which the success or failure of a test is evaluated. proc is one of several Perl subroutines defined within the runtest program. These are primarily used within the testsuite for tests programs exit with expected status values or that timeouts are in fact handled. More control over running tests =============================== You can provide a "processes per node" argument to the run command (typically mpiexec) with either options to "runtests" or environment variables. The two values are -ppnarg=string or MPITEST_PPNARG The string used to specify the number of processes per node. The number of processes to use will be substituted for the %d in the string. For example, export MPITEST_PPNARG="-ppn %d" -ppn=n or MPITEST_PPNMAX The maximum number of processes per node. For example runtests ... -ppn=2 This allows the runtests script to ensure that the value of the processes per node argument does not exceed the total number of processes; some run commands (e.g., aprun on Cray) require that the number of processes per node be no greater than the total number of processes. Note that for most systems it will be important to run the tests multiple times, using this option to ensure that the tests that involve more than one process are run each of the following cases: 1) Multiple MPI processes per chip (likely using shared memory to communicate between processes) 2) MPI processes on separate chips within the same node (also likely using shared memory between processes, but may use a different approach to handle the NUMA nature of this case) 3) MPI processes on separate nodes (likely using the best available interconnect). Note, this depends on the nature of the MPI implementation; these options make it easier to run the necessary cases. If you run only the first case, which is often the default case, you may not effectively test the MPI implementation.