宏观环境
The parallel universe supersedes the mpi universe. The mpi universe eventually will be removed from HTCondor.2. submit范例
#############################################
## submit description file for a parallel universe job
#############################################
universe = parallel
executable = /bin/sleep
arguments = 30
machine_count = 8
log = log
should_transfer_files = IF_NEEDED
when_to_transfer_output = ON_EXIT
queue 3.MPI 范例
######################################
## Example submit description file
## for MPICH 1 MPI
## works with MPICH 1.2.4, 1.2.5 and 1.2.6
######################################
universe = parallel
executable = mp1script
arguments = my_mpich_linked_executable arg1 arg2
machine_count = 4
should_transfer_files = yes
when_to_transfer_output = on_exit
transfer_input_files = my_mpich_linked_executable
queueThe executable is the mp1script script that will have been modified for this MPI application. This script is invoked on each slot or core. The script, in turn, is expected to invoke the MPI application's executable. To know the MPI application's executable, it is the first in the list ofArguments. And, since HTCondor must transfer this executable to the machine where it will run, it is listed with the transfer_input_filescommand, and the file transfer mechanism is enabled with the should_transfer_files command.3. MPI范例2：
Here is the equivalent sample submit description file, but for an Open MPI application:
######################################
## Example submit description file
## for Open MPI
######################################
universe = parallel
executable = openmpiscript
arguments = my_openmpi_linked_executable arg1 arg2
machine_count = 4
should_transfer_files = yes
when_to_transfer_output = on_exit
transfer_input_files = my_openmpi_linked_executable
queue
Most MPI implementations require two system-wide prerequisites. The first prerequisite is the ability to run a command on a remote machine without being prompted for a password. ssh is commonly used. The second prerequisite is an ASCII file containing the list of machines that may utilize ssh. These common prerequisites are implemented in a further script called sshd.sh. sshd.sh generates ssh keys to enable password-less remote execution and starts an sshd daemon. Use of the sshd.sh script requires the definition of two HTCondor configuration variables. Configuration variable CONDOR_SSHDis an absolute path to an implementation of sshd. sshd.sh has been tested with openssh version 3.9, but should work with more recent versions. Configuration variable CONDOR_SSH_KEYGEN points to the corresponding ssh-keygen executable.
Both mp1script and openmpiscript require environmental support. mp1script requires the PATH to the MPICH installation to be set. The shell variableMPdir indicates its proper value. This directory contains the MPICH mpirun executable. Open MPI also uses MPdir.4. vanilla universe环境下运行MPI程序
The vanilla universe may be preferred over the parallel universe for certain parallel applications such as MPI ones. These applications are ones in which the allocated cores need to be within a single slot. The request_cpus command causes a claimed slot to have the required number of CPUs (cores).
There are two ways to ensure that the MPI job can run on any machine that it lands on:
Statically build an MPI library and statically compile the MPI code.
Use CDE to create a directory tree that contains all of the libraries needed to execute the MPI code.
For Linux machines, our experience recommends using CDE, as building static MPI libraries can be difficult. CDE can be found athttp://www.pgbovine.net/cde.html.
Here is a submit description file example assuming that MPI is installed on all machines on which the MPI job may run, or that the code was built using static libraries and a static version of mpirun is available.
############################################################
## submit description file for
## static build of MPI under the vanilla universe
############################################################
universe = vanilla
executable = /path/to/mpirun
request_cpus = 2
arguments = -np 2 my_mpi_linked_executable arg1 arg2 arg3
should_transfer_files = yes
when_to_transfer_output = on_exit
transfer_input_files = my_mpi_linked_executable
queue
If CDE is to be used, then CDE needs to be run first to create the directory tree. On the host machine which has the original program, the command
prompt-> cde mpirun -n 2 my_mpi_linked_executable
creates a directory tree that will contain all libraries needed for the program. By creating a tarball of this directory, the user can package up the executable itself, any files needed for the executable, and all necessary libraries. The following example assumes that the user has created a tarball called cde_my_mpi_linked_executable.tar which contains the directory tree created by CDE.
############################################################
## submit description file for
## MPI under the vanilla universe; CDE used
############################################################
universe = vanilla
executable = cde_script.sh
request_cpus = 2
should_transfer_files = yes
when_to_transfer_output = on_exit
transfer_input_files = cde_my_mpi_linked_executable.tar
transfer_output_files = cde-package/cde-root/path/to/original/directory
queue
The executable is now a specialized shell script tailored to this job. In this example, cde_script.sh contains:
#!/bin/sh
# Untar the CDE package
tar xpf cde_my_mpi_linked_executable.tar
# cd to the subdirectory where I need to run
cd cde-package/cde-root/path/to/original/directory
# Run my command
./mpirun.cde -n 2 ./my_mpi_linked_executable
# Since HTCondor will transfer the contents of this directory
# back upon job completion.
# We do not want the .cde command and the executable transferred back.
# To prevent the transfer, remove both files.
rm -f mpirun.cde
rm -f my_mpi_linked_executable
Any additional input files that will be needed for the executable that are not already in the tarball should be included in the list intransfer_input_files command. The corresponding script should then also be updated to move those files into the directory where the executable will be run.
503
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
