Hands-on batch jobs in Puhti tutorial
The next few exercises take you through submitting your first batch jobs on Puhti. We've used R and HMMER as examples, but the principles are the same for other applications as well. However, please always consult the application specific page if it exists. It may have a tailored template for you.
Get the exercise files
a) Log in to Puhti from a terminal
ssh your-username@puhti.csc.fi
or if you don't have an ssh client available, you can log in via the Puhti web interface using your web browser and open a terminal there.
b) Go to scratch directory and download the exercises file
csc-workspaces
command will show you which projects you're a member of.
csc-workspaces
...
Project: project_20001234 "Great science with HPC tools"
/projappl/project_20001234 16G/54G 38K/100K
/scratch/project_20001234 56G/1.1T 107K/1.0M
------------------------------------------------------------
$HOME
folder!):
and download the input files with wget
:
c) Uncompress the exercises file
Note
Text in italics is not a command and you can choose what to put there (but you need to be consistent later).
A simple batch job script
a) Create a batch job script that prints the compute node on which it is running.
Using nano editor (use whichever editor you like):
Put this into the file.
#!/bin/bash -l
#SBATCH --job-name=print_hostname
#SBATCH --time=00:01:00
#SBATCH --partition=test
#SBATCH --ntasks=1
#SBATCH --mem-per-cpu=1G
#SBATCH --account=project_20001234
sleep 30
echo "This job runs on the host: "; hostname
To exit from the nano editor:
CTRL+O (enter); CTRL+X (confirm save)
Submit the batch script to Puhti
b) Check the job status.
In the following command replace <your username> with your
CSC username - or which ever you used to log in to Puhti. If
you are not sure which it is, you can check it with whoami
or with
this command echo $USER
).
c) What and where did the job print out?
Note
Here you need to replace the JOBID with the ID your job was given.
If the job failed, please check which project you had in the batch script --account=???
Typical reasons for a failed batch in our FAQ
Simple R job
Run a simple R job from a script. The script will fit a straight line through a file containing some x,y value pairs. More info on running R in the CSC environment
a) Set up an interactive batch job and initialize R environment
As we now plan to run an application interactively, we'll ask for an interactive batch job and work there, instead of the login node. The following command will set it up (use the same project as in the batch script above):
Give module load r-env
to initialize the R environment.
- How do you know if it is already loaded?
Go to the directory r-job
, where you have the data (a file called
data.csv
) and launch R:
You should get some introductory text and your prompt should now look like this:
...
Type 'demo()' for some demos, 'help()' for on-line help, or
'help.start()' for an HTML browser interface to help.
Type 'q()' to quit R.
>
b) Create a script to run the same works.
In the "r-job" directory, create an R script file (R
commands to be executed) with the same commands as you pasted
in the R command prompt. Name the file fit.R
You can use e.g. nano
editor, which you need to initialize first
with module load nano
(in the interactive node, in the login node,
it is available without the module load command).
Make sure the file is ok with:
You should see three lines with the R commands.c) Run the script interactively
d) Results
- Did the job succeed?
- What are the fit coefficients?
Simple R job as a batch job
Now run the previous R script as a batch job.
a) Create a batch job script, which will submit the job to the queue.
Copy the serial batch script template from CSC's R-env page into a file called batch.sh
In addition to setting up the computing requirements, this script also resets some additional R environment variables. These won't be necessary, if you run some other application. Always check first if there's a template batch script for your application, and use that as the base for your own script.
For this example, you'll need to make three changes. Replace the --account
and echo "TMPDIR=/scratch/...
lines with your own computing project. And finally, at the end of the script,
replace (myscript.R
) i.e. the R-script to be executed to fit.R
.
b) Submit the batch script with
c) Did the job succeed? Where are the fit constants?
Run tens of R batch jobs as an array job
In this example, we will repeat the previous fitting job for 20 datasets using the array job functionality of SLURM. Note, that for such short jobs it would not make sense to run them as separate batch jobs, but you could loop over them in one job or better inside the R script.
a) Prepare a list of files to process.
Go to the folder named r-array
. Create there a file called
datanames.txt
. This file will contain the names of all those files
that will be used as input to the fitting. Run the following commands to
create it.
b) Write the R script, that will do the fitting.
Go back to the r-array folder, create a script named modelscript.R
and put the following commands to it (you can copy the previous script
and edit that, or start from scratch).
dataname <- commandArgs(trailingOnly = TRUE)
mydata <- read.csv(paste0("data_dir/",dataname))
fit <- lm(y~x,mydata)
write(fit$coefficients,
file=paste0("result_dir/",dataname,"_result.txt"))
The first line will extract from the batch command the name of the dataset to be fitted. The next line reads that data into the variable mydata. Then we fit, like in the previous example, and finally write the coefficients into a file.
c) Create a batch script to submit the job.
Name it R_array.sh
. Copy the contents from the previous example.
Add the following line among the other lines starting #SBATCH:
It will ask SLURM to run an array of 20 jobs. Edit the output and error
files to go to their own directories and files by editing/adding
(%a
will be replaced with the number of the array job):
Change the SLURM partition to be used (--partition=
) from test to small.
Before the line with srun apptainer_wrapper...
, add the following line
You should now have:
datanames.txt
, which has the names of your datafilesmodelscript.R
, which contains the R code to do the fittingR_array.sh
, which is the batch script to submit the job- (and the folders
out, err, data_dir, result_dir
which were there already)
d) run the batch script with
Since you're now running 20 jobs, they might take a moment in the queue. You should get the fit coefficients in separate files in theresult_dir
. Let's now use interactive R to look at the results.
e) Collect the results and plot them.
Note, plotting will work only if you have remote X11 forwarding. Actually, for R, there is even a tailored remote setup using RStudio Server, and you're welcome to use that, but in this tutorial, the key point is to demonstrate the general approach.
In the folder containing analyse.R
start the interactive R shell with
In the R shell that opens, write source("analyse.R")
. This will
run (source) the script contents. The original data was
created by calculating the y values by y=2x + some random noise.
The plot will appear in a separate window.
f) How do the fit coefficients match that?
Batch job with thread parallelization
Some applications can be run in parallel to speed them up. In this example you run the HMMER software to describe and analyze related or similar protein sequence areas both in serial and parallel to see if the jobs speed up.
HMMER uses a database that is already installed, but the protein sequences you want to study need to be copied first to be used as input:
a) Serial HMMER job
Let's first run the job with just one core. Copy one of the old batch scripts to current directory, and change / add the following items in it (or take a look at these examples):
- Output to
out_%j.txt
- error to
err_%j.txt
- run time 10 minutes
- load the
hmmer
-module - remove the R specific environment settings
- run command:
Submit the job with: sbatch your-jobscript-name.sh
Submitting the job echoes the SLURM JOBID number to the screen, but
that is also shown in the output and error filenames
(out_<SLURM_JOBID>.out
). Check if the job is running with
squeue -u <your username>
(use your username)
Or
squeue -j <SLURM_JOBID>
(replace with the JOBID of your job)
Once the job is finished you can check how much memory and time it used:
- Did you reserve a good amount of memory? (not excessively too much, but enough to not be close to memory running out and terminating the job). Another way to get a quick summary of used resources is:
b) Parallel HMMER job
Now, let's try with 4 cores. At this point we'll also switch to using
the environment variable $SLURM_CPUS_PER_TASK
to avoid mistakes
and the need to change that in many places. Add this line to the batch
script:
Note that this only asks the queuing_system for more resources. You need to tell it also to the application you're running (how, depends on the application). Change the run command to:
hmmscan --cpu $SLURM_CPUS_PER_TASK $HMMERDB/Pfam-A.hmm example.fasta > example_$SLURM_CPUS_PER_TASK.result
As you've asked for 4 cpus per task, the environment variable
$SLURM_CPUS_PER_TASK
evaluates to 4 when the script is run, and
you only need to change the number on the #SBATCH line if you want to try other
numbers as well.
Submit the job and check with the sacct
command how long it took to
run the hmmer job and how did the memory usage change and try to answer
these questions:
- Does it make sense to use 4 cores instead of 1?
- Was the memory reservation ok?
- Does it make sense to use more than 4 cores?
- How to speed up the job?
Batch job memory consumption
a) Exceed memory allocation on purpose
Create a new R-script (like in the previous exercise) named mem-test.R
. It
should have the following contents:
dim=10
dim_end=1000
while (dim < dim_end) {
mat <- matrix(rnorm(dim*dim), dim)
print("passed dimension")
print(dim)
dim=dim*2
}
print("all done")
Variable dim
is the dimension of the square matrix, which will be
filled with normal distributed random numbers. The script doubles
dim
until it exceeds dim_end
. Make a new batch script to run your
R-script mem-test.R
. In addition to the resource requests your
script needs to load the R-environment and then run the R-script (as in
previous exercise). Submit the script with sbatch
.
Once the job has been completed (how can you check if it is running or
queuing?), check with sacct
or seff
how much memory was used as
in the previous exercise. If the job completed successfully, increase
the dim_end
variable in your script i.e. make a bigger matrix and
rerun the job. Note also the time it takes to run the job. How does the
time and memory needed by the job depend on the number of elements in
the array?
max(dim) | # of elements | Time [s] (Elapsed) | Memory used (MaxRSS) |
---|---|---|---|
- How big matrix is needed to exceed the default allowed memory of the batch job?
Scaling test for an MPI parallel job
Before running parallel jobs it is important to determine how the job scales. It does not make sense to use many cores, if this does not speed up the job. The speedup depends on the application but also input. In this tutorial we'll use CP2K which can use thousands of cores for certain job types and model systems (but not for this one). Therefore, it is important to test the scaling for each different job type (or model system type). The idea is not to run production simulations, but quick short simulations (i.e. using the actual production system) but only for like 1-5 minutes, which will be enough to reveal the performance.
a) Prepare a CP2k job
First copy the input file to your working directory:
Then create a batch script and submit it with sbatch
Tip
Remember to check the CSC software pages for application specific examples for batch jobs:
For the first job, ask minimal resources (copy the rest of the batch script contents from CSC's CP2k page):
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1
# submit the job
srun cp2k.popt H2O-32.inp > H2O-32_$SLURM_NPROCS.out
b) Run the same job with increasing resources and note performance
After running the job with one core, edit the batch script to use more
cores/mpitasks (e.g. 2,4,8,16, ... this is the --ntasks
or --ntasks-per-node
flag)
and rerun
the job. The output files will be named with the number of cores used to
run them ($SLURM_NPROCS
). Instead of the small partition you can
also use the test
partition. If you ask for more than 40 cores (in Puhti), you
need to switch to large
partition. In multinode jobs, always
limit the number of Nodes, so that the job is not spread onto
more nodes than necessary as it
creates unwanted communication overhead and fragments the allocations on the system.
With the following command you can sum the time spent at different steps for each job.
grep "CPU TIME" H2O-32_1.out | awk '{a+=$5;print a}'
Check with seff JOBID
how much memory the simulation used (compare
sacct
and seff
output!) and fill in the data in the following
table. If your code does not print out either the performance or used
time, you can use the sacct
command (sacct -j <JOBID> -o
elapsed,alloc,maxrss
)
A rule of thumb for acceptable scaling is that when you double the resources, the job should speed up at least 1.5 fold. Ideally it would speed up linearly with resources i.e. 2 fold, so often you're better off running many smaller simulations at the same time, rather than few with very many cores - if you have this option.
The speedup at some core counts may be off the trend. There may be variation due to load on the system or because the code/system does not parallelize well or be able to distribute the computational load for that particular core count. Sometimes rerun helps to sort out an outlier. If you know the code parallelizes well, there is no point to start testing from 1 core, but from where you think the code runs well. You'll notice bad speedup anyway.
# cores | Time [s] (Elapsed) | Speedup | Ideal speedup | Memory used |
---|---|---|---|---|
c) Scaling test results
- How many cores can you use efficiently? (i.e. how far does the job scale)
- How does the required memory depend on the number of cores?
- Why are the elapsed times reported by sacct slightly different to the sum of "CPU TIME" lines?
- Are all nodes similar? Should we limit which resources SLURM may give us?
- If we want to run a different cp2k system do we need to rerun the scaling test?