Advanced MPI scheduling
Most users should submit MPIMessage Passing Interface or distributed memory parallel jobs following the example
given at Running jobs. Simply request a number of
-n and trust the scheduler
to allocate those processes in a way that balances the efficiency of your job
with the overall efficiency of the cluster.
If you want more control over how your job is allocated, then SchedMD's
page on multicore support is a good place to begin. It describes how many of the options to the
command interact to constrain the placement of processes.
You may find this discussion of What exactly is considered a CPU? in Slurm to be useful.
Examples of common MPIMessage Passing Interface scenarios
Few cores, any number of nodes
In addition to the time limit needed for any Slurm job, an MPIMessage Passing Interface job requires that you specify how many MPIMessage Passing Interface processes Slurm should start. The simplest way to do this is with
--ntasks. Since the default memory allocation of 256MB per core is often insufficient, you may also wish to specify how much memory is needed. Using
--ntasks you cannot know in advance how many cores will reside on each node, so you should request memory with
--mem-per-cpu. For example:
--ntasks=15 --mem-per-cpu=3G srun application.exe
This will run 15 MPIMessage Passing Interface processes. The cores could be allocated on one node, on 15 nodes, or on any number in between.
If you have a large parallel job to run, that is, one that can efficiently use 32 cores or more, you should probably request whole nodes. To do so, it helps to know what node types are available at the cluster you are using.
|Graham||32||125 GiB (~3.9 GiB/core)||Some are reserved for whole node jobs.|
|Cedar (Broadwell)||32||125 GiB (~3.9 GiB/core)|
|Cedar (Skylake)||48||187 GiB (~3.9 GiB/core)||Some are reserved for whole node jobs.|
|Niagara||40||188 GiB||Only whole-node requests are possible at Niagara.|
Whole-node jobs are allowed to run on any node. "Some are reserved for whole-node jobs" in the table above indicates that there are nodes on which by-core jobs are forbidden.
A job script requesting whole nodes should look like this:
#!/bin/bash #SBATCH --nodes=2 #SBATCH --ntasks-per-node=32 #SBATCH --mem=0 srun application.exe
#!/bin/bash #SBATCH --nodes=2 #SBATCH --ntasks-per-node=48 #SBATCH --mem=0 srun application.exe
#!/bin/bash #SBATCH --nodes=2 #SBATCH --ntasks-per-node=40 # or 80: Hyperthreading is enabled #SBATCH --mem=0 srun application.exe
--mem=0 is interpreted by Slurm to mean "reserve all the available memory on each node assigned to the job."
Note however that if you need more memory per node than the smallest node provides (e.g. more than 125 GiB at Graham) then you should not use
--mem=0, but request the amount explicitly. Furthermore, some memory on each node is reserved for the operating system, so the largest amount your job can request and still qualify for each node type is as follows:
|Nominal node size||Maximum Slurm request||Where?|
|128 GB = 125 GiB||--mem=128000M||Graham, Cedar Broadwell|
|192 GB = 187 GiB||--mem=192000M||Cedar Skylake|
|256 GB =||--mem=256000M||Graham, Cedar Broadwell|
|512 GB =||--mem=512000M||Graham, Cedar Broadwell|
|1.5 TB =||--mem=1500G||Cedar Broadwell|
|3 TB =||--mem=3000G||Graham, Cedar Broadwell|
--mem=256G instead of
--mem=256000M means the job will not schedule on a so-called 256GB node. The job will not be rejected by Slurm, it will just wait longer that it needs to for a rarer, larger node.
Few cores, single node
If you need less than a full node but need all the cores to be on the same node, then you can request, for example,
--nodes=1 --ntasks-per-node=15 --mem=45G srun application.exe
In this case you could also say
--mem-per-cpu=3G. The advantage of
--mem=45G is that the memory consumed by each individual process doesn't matter, as long as all of them together don’t use more than 45GB. With
--mem-per-cpu=3G, the job will be canceled if any of the processes exceeds 3GB.
Large parallel job, not a multiple of whole nodes
Not every application runs with maximum efficiency on a multiple of 32 (or 40, or 48) cores. Choosing the number of cores to request, and whether or not to request whole nodes, may be a trade-off between running time (or efficient use of the computer) and waiting time (or efficient use of your time). If you want help evaluating these factors, please contact Technical support.
Hybrid jobs: MPIMessage Passing Interface and OpenMP, or MPIMessage Passing Interface and threads
It is important to understand that the number of tasks requested of Slurm is the number of processes that will be started by
srun. So for a hybrid job that will use both MPIMessage Passing Interface processes and OpenMP threads or Posix threads, you should set the MPIMessage Passing Interface process count with
-ntasks-per-node, and set the thread count with
--ntasks=16 --cpus-per-task=4 --mem-per-cpu=3G srun application.exe
In this example a total of 64 cores will be allocated, but only 16 MPIMessage Passing Interface processes (tasks) can and will be initialized. If the application is also OpenMP, then each process will spawn 4 threads, one per core. Each process will be allocated with 12GB of memory. The tasks, with 4 cores each, could be allocated anywhere, from 2 to up to 16 nodes.
--nodes=2 --ntasks-per-node=8 --cpus-per-task=4 --mem=96G srun application.exe
This job is the same size as the last one: 16 tasks (that is, 16 MPIMessage Passing Interface processes), each with 4 threads. The difference here is that we are sure of getting exactly 2 whole nodes. Recall that
--mem requests memory per node, so we use it instead of
--mem-per-cpu for the reason described earlier.
Why srun instead of mpiexec or mpirun?
mpirun is a wrapper that enables communication between processes running on different machines. Modern schedulers already provide many things that
mpirun needs. With Torque/Moab, for example, there is no need to pass to
mpirun the list of nodes on which to run, or the number of processes to launch; this is done automatically by the scheduler. With Slurm, the task affinity is also resolved by the scheduler, so there is no need to specify things like
mpirun --map-by node:pe=4 -n 16 application.exe
As implied in the examples above,
srun application.exe will automatically distribute the processes to precisely the resources allocated to the job.
In programming terminology,
srun is higher level of abstraction than
mpirun. Anything that can be done with
mpirun can be done with
srun, and more. It is the tool in Slurm to distribute any kind of computations. It replaces Torque’s
pbsdsh, for example, and much more. Think of
srun as the SLURM "all-around parallel-tasks distributor"; once a particular set of resources is allocated, the nature of your application doesn't matter (MPIMessage Passing Interface, OpenMP, hybrid, serial farming, pipelining, multi-program, etc.), you just have to
Also, as you would expect,
srun is fully coupled to Slurm. When you
srun an application, a "job step" is started, the environment variables
SLURM_PROCID are initialized correctly, and correct accounting information is recorded.
For an example of some differences between
mpiexec, see this discussion on the Open MPIMessage Passing Interface support forum. Better performance might be achievable with
mpiexec than with
srun under certain circumstances, but using
srun minimizes the risk that there will be a mismatch between the resources allocated by Slurm and those used by Open MPIMessage Passing Interface.