15 KiB
| title | slug | date | categories | author | cc_license | description | aliases | |
|---|---|---|---|---|---|---|---|---|
| VSC/Slurm/Spack Tips and Tricks | vsc-slurm-spack-tips-and-tricks | 2022-05-18 | cheatsheet | Lukas Winkler | true | An assorted list of tricks for using the Vienna Scientific Cluster |
|
This is not official documentation for the Vienna Scientific Cluster. For this check the VSC Wiki. Instead, this is my personal cheat sheet of things that are not well documented elsewhere. Also while the content is focused on the VSC, most of the things mentioned here also apply to similar setups that use Slurm at other universities.
Basics
Always request an interactive session when running anything using a non-trivial amount of CPU power!
Quick interactive session
➜ salloc --ntasks=2 --mem=2G --time=01:00:00
Don't forget to then connect to the node you get assigned:
➜ ssh n1234-567
Storage
official docs:
$HOME is limited to 100 GB and storing/compiling code. Anything else should be stored at $DATA.
Quota
The file size and number of files is limited by group. The current status can be read using
➜ mmlsquota --block-size auto -j data_fs00000 data
for $DATA and
➜ mmlsquota --block-size auto -j home_fs00000 home
for $HOME where 00000 is the ID of the own project (accessible using groups)
Job scripts
Basic Template
Your job script is a regular bash script (.sh file). In addition, you can specify options to sbatch in the beginning of your file:
#!/bin/bash
#SBATCH --job-name=somename
#SBATCH --mail-type=ALL
#SBATCH --mail-user=yourmail@example.com
--job-name: A short name for this job. It is often displayed truncated after a few characters.--mail-type=ALL: notify on all events per E-Mail
In these cases --long-option=value and --long-option value are equivalent.
Single Core job
Only specify --ntask=1 and the amount of memory you need.
#SBATCH --ntasks=1 # (also -n 1)
#SBATCH --mem 2G
More sbatch options
All options can be found in the slurm documentation. A few useful ones are:
--nodes,-N: request X nodes--ntasks--time: limit the run time, e.g.10:00for 10 hours--mem: memory limit, e.g.20G--output,-o: where to store the output of the executed script--dependency=afterany:1234: only run job after job with ID 1234 has terminated
Useful Environment Variables
$SLURM_JOB_NAME$SLURM_NODELIST$SLURM_NNODES$SLURM_NPROCS
Especially the latter can be used e.g. for running MPI programs with the requested number of CPU cores:
➜ mpiexec -np $SLURM_NPROCS ./program
Submitting Jobs
A job script can be submitted using
➜ sbatch jobfile.sh # you can also add sbatch options here
Just like in regular shell scripts, you can pass arguments to jobfile.sh like this
➜ sbatch jobfile.sh somevalue
and then access somevalue as $1 in your script. This way multiple similar jobs can be submitted without needing to edit the jobscript.
Queue
The current status of jobs in the Queue can be seen using squeue.
➜ squeue -u username
Especially useful is the estimated start time of a scheduled job:
➜ squeue -u username --start
A lot more information about scheduling including the calculated priority of jobs can be found using sprio
➜ sprio -u username
This will also show the reason why the job is still queued for which an explanation can be found in the slurm documentation or the VSC wiki.
Details about past Jobs (like maximum memory usage), can be found using sacct. You can manually specify the needed columns or display most of them using --long
➜ sacct -j 2052157 --long
Advanced Slurm features
QoS, accounts and partitions
Depending on access to private nodes, you might have access to many different QoS (Quality of Service), accounts and partitions.
On VSC you can get an overview over your account with sqos (this is also shown on login):
➜ sqos -acc # this only works on VSC
If you want to a different account or QoS than your default (e.g. if you want to access private nodes or GPU nodes), you can specify them with --qos and --acccount in salloc, sbatch or your job script.
You can also get an overview over all available partitions with sinfo and specify one explicitly with --partition.
If you want to get a quick overview over the QoS at VSC and their current usage, you can use sqos.
Array Jobs
Sometimes you might want to submit a larger number of similar jobs. This can be easily achieved using array jobs and the --array argument. With this, your job will be submitted multiple times with a different task ID that can be used from the $SLURM_ARRAY_TASK_ID environment variable.
#SBATCH --array=0-26
./your_program $SLURM_ARRAY_TASK_ID
{{< alert type="warning" >}} Keep in mind that each individual job should not be too small (more than just a few minutes) as otherwise the computational overhead of scheduling the job and starting it will not be worth it. In these cases using one job that runs the program in a loop will be more efficient. {{< /alert >}}
SSH login via login.univie.ac.at
official docs (but we are using the more modern ProxyJump instead of Agent forwarding as this way we don't have to trust the intermediate server with our private key)
Access to VSC is only possible from IP addresses of the partner universities. If you are from the University of Vienna and don't want to use the VPN, an SSH tunnel via login.univie.ac.at is an alternative.
To connect to the login server, the easiest thing is to put the config for the host in your ~/.ssh/config (create it, if it doesn't yet exist).
Host loginUnivie
HostName login.univie.ac.at
User testuser12 # replace with your username
# the following are needed if you are using OpenSSH 8.8 or newer
# and the login server isn't yet updated to a never version
HostkeyAlgorithms +ssh-rsa
PubkeyAcceptedAlgorithms +ssh-rsa
This way you should now be able to test connecting to the login server using
➜ ssh loginUnivie
Then you can add another entry to ~/.ssh/config on your computer for VSC that uses ProxyJump to connect via the loginUnivie entry we just created.
Host vsc5
Hostname vsc5.vsc.ac.at
User vscuser
ProxyJump loginUnivie
# Port 27 # (only use if you are using ssh keys)
➜ ssh vsc5
Spack Modules
(official docs, that this guide builds on. More useful tips can be found in the spack documentation)
Software that is needed can be loaded via modules. The easiest way to find the right module for the current processor architecture, is directly querying spack, which is used to provide all compiled libraries and applications. There should never be a need to run module directly and doing so might accidentally pick libraries that are not intended for the current processor architecture.
Finding the right module
The easiest way is using spack find.
➜ spack find cmake
If you get a long output, you can ignore everything above the ==> N installed package(s) line as it is unrelated to your current query. In case this only returns one module that fits your requirements, you can directly replace spack find with spack load to load this module.
But most of the time, you will find multiple modules which differ in their properties (and spack load will fail if the query resolves to more than one package):
➜ spack find cmake
==> 4 installed packages
-- linux-almalinux8-zen / gcc@8.5.0 -----------------------------
cmake@3.21.4
-- linux-almalinux8-zen2 / intel@2021.5.0 -----------------------
cmake@3.21.4
-- linux-almalinux8-zen3 / aocc@3.2.0 ---------------------------
cmake@3.21.4
-- linux-almalinux8-zen3 / gcc@11.2.0 ---------------------------
cmake@3.21.4
The most important property is the version and it is denoted with an @ sign. Another property is the compiler the program or library was compiled with and it can be separated with a %.
So if you want to load e.g. cmake version 3.x.x compiled with gcc version 11, you could directly search for it and subsequently load it.
➜ spack find cmake@3%gcc@11
➜ spack load cmake@3%gcc@11
This way if another minor update of cmake is released, your command will load it. If you don't like this, check the next section.
Sometimes there are also multiple variants of the same module. spack info modulename can give you an overview over all of them, but that doesn't mean that all combinations of variants/compilers/versions are offered at VSC. If you are for example interested in the hdf5 library with MPI support, you can search for the following (-v gives you the exact properties of each module):
➜ spack find -v hdf5 +mpi
"Locking" modules
If you dislike the fact that spack load queries don't resolve to specific packages, but just filters that describe the properties you want or prefer exactly specifying the version of a package for reproducibility, you can find the hash of package using spack find -l and can then use /hash to always refer to this exact package:
➜ spack find -l gsl
==> 1 installed package
-- linux-almalinux8-zen3 / gcc@11.2.0 ---------------------------
4rhrhm3 gsl@2.7
$ spack load /4rhrhm3
Find currently loaded modules
# List all currently loaded packages
➜ spack find --loaded
# Unload all currently loaded packages
➜ spack unload --all
Avoiding broken programs due to loaded dependencies
Loading a spack module not just loads the specified module, but also all dependencies of this module. With some modules like openmpi that dependency tree can be quite large.
➜ spack find -d openmpi%gcc
-- linux-almalinux8-zen3 / gcc@11.2.0 ---------------------------
openmpi@4.1.4
hwloc@2.6.0
libpciaccess@0.16
libxml2@2.9.12
libiconv@1.16
xz@5.2.5
zlib@1.2.11
ncurses@6.2
libevent@2.1.8
openssl@1.1.1l
numactl@2.0.14
openssh@8.7p1
libedit@3.1-20210216
pmix@3.2.1
slurm@22-05-2-1
curl@7.79.0
glib@2.70.0
gettext@0.21
bzip2@1.0.8
tar@1.34
libffi@3.3
pcre@8.44
perl@5.34.0
berkeley-db@18.1.40
gdbm@1.19
readline@8.1
python@3.8.12
expat@2.4.1
libbsd@0.11.3
libmd@1.0.3
sqlite@3.36.0
util-linux-uuid@2.36.2
json-c@0.15
lz4@1.9.3
munge@0.5.14
libgcrypt@1.9.3
libgpg-error@1.42
ucx@1.12.1
And loading module like openssl or ncurses from spack means that programs that depend on those libraries, but the versions provided by the base operating system, will crash.
➜ spack load openmpi%gcc
➜ nano somefile.txt
Segmentation fault (core dumped)
➜ htop
Segmentation fault (core dumped)
One can avoid this by unloading the affected modules afterwards.
➜ spack unload ncurses
➜ spack unload openssl
But in many cases one doesn't need all dependency modules and is really just interested in e.g. openmpi itself. Therefore, one can ignore the dependencies with --only package.
# doesn't affect non-openmpi programs
➜ spack load --only package openmpi%gcc
Comparing modules
Sometimes two packages look exactly the same:
➜ spack find -vl fftw
-- linux-almalinux8-zen2 / intel@2021.5.0 -----------------------
mmgor5w fftw@3.3.10+mpi+openmp~pfft_patches precision=double,float cy5tkce fftw@3.3.10+mpi+openmp~pfft_patches precision=double,float
Then you can use spack diff to
➜ spack diff /mmgor5w /cy5tkce
--- fftw@3.3.10/mmgor5w3daiwtsdbyl4dfhjsueaciry2
+++ fftw@3.3.10/cy5tkcetpgx35rok2lqfi3d66rjptkva
@@ depends_on @@
- fftw intel-oneapi-mpi build
+ fftw openmpi build
[...]
Therefore, we know that in this example the first package depends on intel-oneapi-mpi and the second one on openmpi.
Debugging modules
Sometimes one needs to know what spack load somepackage does exactly (e.g. because a library is still not found even though you loaded the module). Adding --sh to spack load prints out all commands that would be executed during the module load allowing you to understand what is going on.
➜ spack load --sh cmake%gcc@8
export ACLOCAL_PATH=[...];
export CMAKE_PREFIX_PATH=[...];
export CPATH=[...];
export LD_LIBRARY_PATH=[...];
export LIBRARY_PATH=[...];
export MANPATH=[...];
export PATH=[...];
export PKG_CONFIG_PATH=[...];
export SPACK_LOADED_HASHES=[...];
Commonly used modules
This is a list of modules I commonly use. While it might not be directly usable for other people and will go out of date quickly, it might still serve as a good starting point.
spack load --only package openmpi@4%gcc@11.2/rpec5sw
spack load --only package fftw%gcc@11.2
spack load --only package libtool%gcc@11.2 # GNU Autotools
spack load --only package hdf5%gcc@11.2/uglkavv # hdf5%gcc@11.2 +mpi
spack load --only package numactl%gcc@11.2
spack load --only package metis%gcc@11.2
spack load --only package intel-tbb%gcc@11.2
spack load --only package gsl%gcc@11.2
spack load --only package cmake%gcc@11.2
# spack load --only package gcc@11.2
module load gcc/11.2.0-gcc-11.2.0-5i4t2bo
spack load --only package python@3.11%gcc@11