HTCondor’s parallel universe supports jobs that span multiple machines, where the multiple processes within a job must be running concurrently on these multiple machines, perhaps communicating with each other. The parallel universe provides machine scheduling, but does not enforce a particular programming paradigm for the underlying applications. Thus, parallel universe jobs may run under various MPI implementations as well as under other programming environments.
The parallel universe supersedes the mpi universe. The mpi universe eventually will be removed from HTCondor.
Parallel universe jobs are submitted from the machine running the dedicated scheduler. The dedicated scheduler matches and claims a fixed number of machines (slots) for the parallel universe job, and when a sufficient number of machines are claimed, the parallel job is started on each claimed slot.
Each invocation of condor_submit assigns a single ClusterId for what is considered the single parallel job submitted. The machine_count submit command identifies how many machines (slots) are to be allocated. Each instance of the queue submit command acquires and claims the number of slots specified by machine_count. Each of these slots shares a common job ClassAd and will have the same ProcId job ClassAd attribute value.
Once the correct number of machines are claimed, the executable is started at more or less the same time on all machines. If desired, a monotonically increasing integer value that starts at 0 may be provided to each of these machines. The macro $(Node) is similar to the MPI rank construct. This macro may be used within the submit description file in either the arguments or environment command. Thus, as the executable runs, it may discover its own $(Node) value.
Node 0 has special meaning and consequences for the parallel job. The completion of a parallel job is implied and taken to be when the Node 0 executable exits. All other nodes that are part of the parallel job and that have not yet exited on their own are killed. This default behavior may be altered by placing the line
in the submit description file. It causes HTCondor to wait until every node in the parallel job has completed to consider the job finished.
To run parallel universe jobs, HTCondor must be configured such that machines running parallel jobs are dedicated. Note that dedicated has a very specific meaning in HTCondor: while dedicated machines can run serial jobs, they prefer to run parallel jobs, and dedicated machines never preempt a parallel job once it starts running.
A machine becomes a dedicated machine when an administrator configures it to accept parallel jobs from one specific dedicated scheduler. Note the difference between parallel and serial jobs. While any scheduler in a pool can send serial jobs to any machine, only the designated dedicated scheduler may send parallel universe jobs to a dedicated machine. Dedicated machines must be specially configured. See section 3.14.9 for a description of the necessary configuration, as well as examples. Usually, a single dedicated scheduler is configured for a pool which can run parallel universe jobs, and this condor_schedd daemon becomes the single machine from which parallel universe jobs are submitted.
The following command line will list the execute machines in the local pool which have been configured to use a dedicated scheduler, also printing the name of that dedicated scheduler. In order to run parallel jobs, this name will be defined to be the string "DedicatedScheduler@", prepended to the name of the scheduler host.
If this command emits no lines of output, then then pool is not correctly configured to run parallel jobs. Make sure that the name of the scheduler is correct. The string after the @ sign should match the name of the condor_schedd daemon, as returned by the command
Here is a submit description file for a parallel universe job example that is as simple as possible:
This job specifies the universe as parallel, letting HTCondor know that dedicated resources are required. The machine_count command identifies that eight machines are required for this job.
Because no requirements are specified, the dedicated scheduler claims eight machines with the same architecture and operating system as the submit machine. When all the machines are ready, it invokes the /bin/sleep command, with a command line argument of 30 on each of the eight machines more or less simultaneously. Job events are written to the log specified in the log command.
The file transfer mechanism is enabled for this parallel job, such that if any of the eight claimed execute machines does not share a file system with the submit machine, HTCondor will correctly transfer the executable. This /bin/sleep example implies that the submit machine is running a Unix operating system, and the default assumption for submission from a Unix machine would be that there is a shared file system.
Assume that the pool contains Linux machines installed with either a RedHat or an Ubuntu operating system. If the job should run only on RedHat platforms, the requirements expression may specify this:
The machine selection may be further narrowed, instead using the OpSysAndVer attribute.
The $(Node) macro is expanded to values of 0-3 as the job instances are about to be started. This assigns unique names to the input and output files to be transferred or accessed from the shared file system. The $(Node) value is fixed for the entire length of the job.
Sometimes one machine’s part in a parallel job will have specialized needs. These can be handled with a Requirements submit command that also specifies the number of needed machines.
The dedicated scheduler acquires and claims four machines. All four share the same value of ClusterId, as this value is associated with this single parallel job. The existence of a second queue command causes a total of two ProcId values to be assigned for this parallel job. The ProcId values are assigned based on ordering within the submit description file. Value 0 will be assigned for the single executable that must be executed on machine1@example.com, and the value 1 will be assigned for the other three that must be executed elsewhere.
If the parallel program has a structure that benefits from running on multiple cores within the same slot, multi-core slots may be specified.
This parallel job causes the scheduler to match and claim two machines, where each of the machines (slots) has eight cores. The parallel job is assigned a single ClusterId and a single ProcId, meaning that there is a single job ClassAd for this job.
The executable, foo.sh, is started at the same time on a single core within each of the two machines (slots). It is presumed that the executable will take care of invoking processes that are to run on the other seven CPUs (cores) associated with the slot.
Potentially fewer machines are impacted with this specification, as compared with the request that contains
The interaction of the eight cores within the single slot may be advantageous with respect to communication delay or memory access. But, 8-core slots must be available within the pool.
MPI applications use a single executable, invoked on one or more machines (slots), executing in parallel. The various implementations of MPI such as Open MPI and MPICH require further framework. HTCondor supports this necessary framework through a user-modified script. This implementation-dependent script becomes the HTCondor executable. The script sets up the framework, and then it invokes the MPI application’s executable.
The scripts are located in the $(RELEASE_DIR)/etc/examples directory. The script for the Open MPI implementation is openmpiscript. The scripts for MPICH implementations are mp1script and mp2script. An MPICH3 script is not available at this time. These scripts rely on running ssh for communication between the nodes of the MPI application. The ssh daemon on Unix platforms restricts connections to the approved shells listed in the /etc/shells file.
Here is a sample submit description file for an MPICH MPI application:
The 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 of arguments. And, since HTCondor must transfer this executable to the machine where it will run, it is listed with the transfer_input_files command, and the file transfer mechanism is enabled with the should_transfer_files command.
Here is the equivalent sample submit description file, but for an Open MPI application:
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_SSHD is 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.
mp1script and mp2script require the PATH to the MPICH installation to be set. The variable MPDIR may be modified in the scripts to indicate its proper value. This directory contains the MPICH mpirun executable.
openmpiscript also requires the PATH to the Open MPI installation. Either the variable MPDIR can be set manually in the script, or the administrator can define MPDIR using the configuration variable OPENMPI_INSTALL_PATH . When using Open MPI on a multi-machine HTCondor cluster, the administrator may also want to consider tweaking the OPENMPI_EXCLUDE_NETWORK_INTERFACES configuration variable as well as set MOUNT_UNDER_SCRATCH = /tmp.
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:
For Linux machines, our experience recommends using CDE, as building static MPI libraries can be difficult. CDE can be found at http://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.
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
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.
The executable is now a specialized shell script tailored to this job. In this example, cde_script.sh contains:
Any additional input files that will be needed for the executable that are not already in the tarball should be included in the list in transfer_input_files command. The corresponding script should then also be updated to move those files into the directory where the executable will be run.