I've got an mpi job I run in slurm using an sbatch script which looks something like:
# request 384 processors across 16 nodes for exclusive use:
#SBATCH --exclusive
#SBATCH --ntasks-per-node=24
#SBATCH -n 384
#SBATCH -N 16
#SBATCH --time 3-00:00:00
mpirun myprog
I want to monitor the memory/cpu usage and some other behaviour of the "myprog" processes. I've written a simple script (call it "monitor") which can do this, but I'm stumped on how to use sbatch to run ONE copy of it on each allocated node, at the same time as "myprog".
I think I need to modify the above to something like:
...
srun monitor
mpirun myprog
But I'm confused about whether a) that means "monitor" will run in the background and b) how I can control where "monitor" runs.
To have monitor run 'in the background', so actually the srun is non-blocking and the subsequent mpirun command can start, you simply need to add an ampersand (&) at the end.
To make sure that program runs on the 'master node' of the allocation, just remove the srun command.
If you need that program to run on a specific node, use the -n1 --nodelist option (you probably first need to get the list of all allocated nodes first.) You should also consider using the --overcommit option of srun to avoid dedicating a full CPU to your monitoring program which I assume is not CPU-bound.
Related
I want to run two programs using mpi in parallel in the same job script. In SLURM I would usually just write a script for sbatch (shortened):
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=4
mpirun program1 &
mpirun program2
This works fine.
The two programs will internally communicate with each other and coordinate execution. So overcommiting is fine. Moreover, they require each other and cannot run as stand-alone in the present configuration.
However, if I want to extend this to several nodes, e.g.
#SBATCH --nodes=2
#SBATCH --ntasks-per-node=2
SLURM does not start the first job in the background. Instead, it starts in the foreground, fails because it does not find the second step and the second then also fails -- because it does not find the first.
I am a bit at a loss here because that is the suggested solution (e.g. Run a "monitor" task alongside mpi task in SLURM) to similar problems and I do not see a reason why this should not work over several nodes. Indeed it does, for instance on PBS.
You can run your Multiple Program Multiple Data (MPMD) application like this:
mpirun -np x program1 : -np y program2
How do I run setup code in a SLURM sbatch script? Can I just use two srun lines?
Are these two srun lines guaranteed to run on the same node, without cleanup inbetween?
#!/bin/bash
# Parameters
#SBATCH ...
# setup
srun cp /nfs/data $TMPDIR
# job
srun a.out $TMPDIR
The srun command will start as many instances of the command as specified with the --ntasks parameter. It is typically used with MPI programs and programs that run embarrassingly parallel workloads.
A command like srun cp ... only makes sense in the case multiple nodes are requested and only one task is running per node, so with for instance --nodes=N or --ntasks=N --ntasks-per-node=1 or a similar combination. It can be used to copy files from a network filesystem to a local filesystem.
If there is only one node and multiple tasks, the srun could cause problems by concurrently trying to write to the same file.
If there is only one task, then the srun are not really needed (except if you want to use sstat to monitor them).
In any case, consecutive srun's are run on the same sets of nodes without cleaning.
I am new to Slurm and I also found the related questions about this topic. However, I am still confused about several points of how to use srun. According to the official document, srun will typically first allocate resources and then run the parallel jobs. For example, I want to run 20 tasks and if I submit my job based on the following script, I am not sure how many tasks are created. Because sbatch only takes care of allocating resources instead of executing program.
#!/bin/sh
#SBATCH -n 20
#SBATCH --mpi=pmi2
#SBATCH -o myoutputfile.txt
module load mpi/mpich-x86_64
mpirun mpiprogram < inputfile.txt
If I am trying to run sequential program like the following, I am not whether there will be a difference or not. For example, I can simply remove the srun command in this script. What will happen?
#!/bin/sh
#SBATCH -n 1
#SBATCH -N 1
srun tar zxf julia-0.3.11.tar.gz
echo "prefix=/software/julia-0.3.11" > julia/Make.user
cd julia
srun make
The first example will spawn 20 tasks ; sbatch will request 20 CPUs and also set up the environment so that mpirun knows how many CPUs were requested for the job. mpirun will then spawn as many processes as were allocated (provided that OpenMPI was compiled with Slurm support).
The #SBATCH --mpi=pmi2 part is meant for srun so it will have no effect if srun is not called in the submission script.
In the second example, there will be no difference in the number of processes spawned as only one is needed. But, with srun, the output of sstat will be more reliable, the management of signals will be more precise, and the buffering of the output will be more controlled (via the srun command line options).
If you request multiple tasks, srun will instantiate that many processes. It can be an MPI program, or a sequential program that adapts its behaviour based on the SLURM_PROC_ID environment variable.
Also you can run multiple srun in the same submission script. Each instance of srun (called a "step") is then accounted separately in the accounting (sacct).
Finally, srun can use a subset of the allocation and organise the micro-scheduling of many small tasks in a single job (see the example in the srun manpage).
I have access to a large GPU cluster (20+ nodes, 8 GPUs per node) and I want to launch a task several times on n GPUs (1 per GPU, n > 8) within one single batch without booking full nodes with the --exclusive flag.
I managed to pre-allocate the resources (see below), but I struggle very hard with launching the task several times within the job. Specifically, my log shows no value for the CUDA_VISIBLE_DEVICES variable.
I know how to do this operation on fully booked nodes with the --nodes and --gres flags. In this situation, I use --nodes=1 --gres=gpu:1 for each srun. However, this solution does not work for the present question, the job hangs indefinitely.
In the MWE below, I have a job asking for 16 gpus (--ntasks and --gpus-per-task). The jobs is composed of 28 tasks which are launched with the srun command.
#!/usr/bin/env bash
#SBATCH --job-name=somename
#SBATCH --partition=gpu
#SBATCH --nodes=1-10
#SBATCH --ntasks=16
#SBATCH --gpus-per-task=1
for i in {1..28}
do
srun echo $(hostname) $CUDA_VISIBLE_DEVICES &
done
wait
The output of this script should look like this:
nodeA 1
nodeR 2
...
However, this is what I got:
nodeA
nodeR
...
When you write
srun echo $(hostname) $CUDA_VISIBLE_DEVICES &
the expansion of the $CUDA_VISIBLE_DEVICES variable will be performed on the master node of the allocation (where the script is run) rather than on the node targeted by srun. You should escape the $:
srun echo $(hostname) \$CUDA_VISIBLE_DEVICES &
By the way, the --gpus-per-task= appeared in the sbatch manpage in the 19.05 version. When you use it with an earlier option, I am not sure how it goes.
I've installed Slurm on a 2-node cluster. Both nodes are compute nodes, one is the controller also. I am able to successfully run srun with multiple jobs at once. I am running GPU jobs and have confirmed I can get multiple jobs running on multiple GPUs with srun, up to the number of GPUs in the systems.
However, when I try running sbatch with the same test file, it will only run one batch job, and it only runs on the compute node which is also the controller. The others fail, with an ExitCode of 1:0 in the sacct summary. If I try forcing it to run on the compute node that's not the controller, it won't run and shows the 1:0 exit code. However, just using srun will run on any compute node.
I've made sure the /etc/slurm/slurm.conf files are correct with the specs of the machines. Here is the sbatch .job file I am using:
#!/bin/bash
#SBATCH --job-name=tf_test1
#SBATCH --output=/storage/test.out
#SBATCH --error=/storage/test.err
#SBATCH --ntasks=2
#SBATCH --cpus-per-task=1
#SBATCH --mem-per-cpu=2000
##SBATCH --mem=10gb
#SBATCH --gres=gpu:1
~/anaconda3/bin/python /storage/tf_test.py
Maybe there is some limitation with sbatch I don't know about?
sbatch creates a job allocation and launches what is called the 'batch step'.
If you aren't familiar with what a job step is, I recommend this page: https://slurm.schedmd.com/quickstart.html
The batch step runs the script passed to it from sbatch. The only way to launch additional job steps is to invoke srun inside the batch step. In your case, it would be
srun ~/anaconda3/bin/python /storage/tf_test.py
This will create a job step running tf_test.py on each task in the allocation. Note that while the command is the same as when you run srun directly, it detects that is inside an allocation via environment variables from sbatch. You can split up the allocation into multiple job steps by running srun with flags like -n[num tasks] instead. ie
#!/bin/bash
#SBATCH --ntasks=2
srun --ntasks=1 something.py
srun --ntasks=1 somethingelse.py
I don't know if you're having any other problems because you didn't post any other error messages or logs.
If using srun on the second node works and using sbatch with the submission script you mention fails without any output written, the most probable reason would be that /storage does not exist, or is not writable by the user, on the second node.
The slurmd logs on the second node should be explicit about this. The default location is /var/log/slurm/slurmd.log, but check the output of scontrol show config| grep Log for definitive information.
Another probable cause that lead to the same behaviour would be that the user is not defined or has a different UID on the second node (but then srun would fail too)
#damienfrancois answer was closest and maybe even correct. After making sure the /storage location was available on all nodes, things run with sbatch. The biggest issue was the /storage location is shared via NFS, but it was read-only for the compute nodes. This had to be changed in /etc/exports to look more like:
/storage *(rw,sync,no_root_squash)
Before it was ro...
The job file I have that works is also a bit different. Here is the current .job file:
#!/bin/bash
#SBATCH -N 1 # nodes requested
#SBATCH --job-name=test
#SBATCH --output=/storage/test.out
#SBATCH --error=/storage/test.err
#SBATCH --time=2-00:00
#SBATCH --mem=36000
#SBATCH --qos=normal
#SBATCH --mail-type=ALL
#SBATCH --mail-user=$USER#nothing.com
#SBATCH --gres=gpu
srun ~/anaconda3/bin/python /storage/tf_test.py