Running Python on the ORC Clusters
[!NOTE] To run Python jobs or build Python Virtual Environments that will run across all the nodes on Hopper, use the Python modules under the GCC/10 compiler:
``` module load gnu10 module load python/<version> ## the default python is version 3.9.9 ```
The examples below will be based on the Hopper cluster. Slight modifications in the Slurm scripts will be necessary to run them on Argo.
Python Versions
To see the available version of Python, run the command
ml spider python
This will list all the available versions of Python that are installed on the cluster and include all the different builds.
------------------------------------------------------------------------------------------------
python:
------------------------------------------------------------------------------------------------
Versions:
python/2.7.18-z2
python/2.7.18-z4
python/3.7.4-rg
python/3.7.6-iu
python/3.7.6-ks
python/3.7.7-intel
python/3.8.6-ff
python/3.8.6-pi
python/3.8.6-kg
python/3.8.6-rp
python/3.8.6-vw
python/3.8.6-ye
python/3.8.6-p2
python/3.8.6-4q
python/3.9.7-intel
python/3.9.9-jh
Other possible modules matches:
intel-python intelpython
------------------------------------------------------------------------------------------------
To find other possible module matches execute:
$ module -r spider '.*python.*'
------------------------------------------------------------------------------------------------
module load gnu10
module avail python
-------------------------------------------- GNU-10.3.0 ---------------------------------------------
python/3.8.6-pi python/3.9.9-jh (D)
Where:
D: Default Module
Running
module load python
module load python/<version>
Running a Python Job
Interactively on a CPU
Python jobs should not be run directly on the head nodes. The preferred method, even if you're testing
a small job, is to start a debug session directly on a compute node using the debug partition and then test your script or, for short jobs,
run it directly from the node. The default time limit on the debug partition is 1 hour. To get more information on the available partitions, resources, and limits on the node use the sinfo
command.
To connect directly to a compute node and use the debug partition, use the salloc
command together with additional Slurm parameters
salloc -p debug -n 1 --cpus-per-task=12 --mem=15GB
salloc -p normal -n 1 --cpus-per-task=12 --mem=15GB -t 0-02:00:00
salloc: Granted job allocation
salloc: Waiting for resource configuration
salloc: Nodes hop065 are ready for job
[user@hop065 ~]$
[user@hop065 ~]$ module list
Currently Loaded Modules:
1) use.own 3) prun/2.0 5) gnu10/10.3.0-ya 7) sqlite/3.37.1-6s 9) openmpi/4.1.2-4a
2) autotools 4) hosts/hopper 6) zlib/1.2.11-2y 8) tcl/8.6.11-d4 10) python/3.9.9-jh
Inactive Modules:
1) openmpi4
[user@hop065 ~]$ python
Python 3.9.9 (main, Mar 25 2022, 16:08:31)
[GCC 10.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>>
$ python myscript.py
The interactive session will persist until you type the 'exit' command as shown below:
$ exit
exit
salloc: Relinquishing job allocation
Interactively on a GPU
In a similar manner, you can start an interactive session on a GPU node with
salloc -p gpuq -q gpu --ntasks-per-node=1 --gres=gpu:A100.40gb:1 -t 0-01:00:00
Using a Slurm Submission Script
Once your tests are done and you're ready to run longer Python jobs, you should now switch to using
the batch submission with Slurm. To do this, you write a Slurm script setting the different parameters
for your job, loading the necessary modules, and executing your Python script which is then submitted to the
selected queue from where it will run your job. Below is an example Slurm script (run.slurm
):
#!/bin/bash
#SBATCH --partition=normal # will run on any cpus in the 'normal' partition
#SBATCH --job-name=python-cpu
## Separate output and error messages into 2 files.
## NOTE: %u=userID, %x=jobName, %N=nodeID, %j=jobID
#SBATCH --output=/scratch/%u/%x-%N-%j.out # Output file
#SBATCH --error=/scratch/%u/%x-%N-%j.err # Error file
#SBATCH --nodes=1
#SBATCH --cpus-per-task=1 # up to 48 per node
#SBATCH --mem-per-cpu=3500M # memory per CORE; maximum is 180GB per node
#SBATCH --export=ALL
#SBATCH --time=0-01:00:00 # set to 1hr; please choose carefully
set echo
umask 0027
module load gnu10
module load python # load the recommended Python version
python myscript.py # execute your Python script
If you need GPU nodes for your Python job, you would change the partition information to the gpuq
and set the number of GPU nodes needed.
#!/bin/bash
#SBATCH --partition=gpuq # the DGX only belongs in the 'gpu' partition
#SBATCH --qos=gpu # need to select 'gpu' QoS
#SBATCH --job-name=python-gpu
## Separate output and error messages into 2 files.
## NOTE: %u=userID, %x=jobName, %N=nodeID, %j=jobID
#SBATCH --output=/scratch/%u/%x-%N-%j.out # Output file
#SBATCH --error=/scratch/%u/%x-%N-%j.err # Error file
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1 # up to 128;
#SBATCH --gres=gpu:A100.80gb:1 # up to 8; only request what you need
#SBATCH --mem-per-cpu=3500M # memory per CORE; total memory is 1 TB (1,000,000 MB)
#SBATCH --export=ALL
#SBATCH --time=0-01:00:00 # set to 1hr; please choose carefully
set echo
umask 0027
# to see ID and state of GPUs assigned
nvidia-smi
module load gnu10
module load python
python myscript.py
Please use the scratch space to submit your job's Slurm script with
sbatch run.slurm
cd /scratch/UserID
command to change directories(replace 'UserId' with your GMU GMUnetID). Please note that scratch directories have no space limit and data in /scratch gets purged 90 days from the date of creation, so make sure to move your files to a safe place before the purge.
To copy files directly from scratch to your project space you can use the cp
command to create a copy of the contents of the file or directory specified by the SourceFile or SourceDirectory parameters into the file or directory specified by the TargetFile or TargetDirectory parameters. The cp
command also copies entire directories into other directories if you specify the -r or -R flags.
The command below copies entire files from the scratch space to your project space (" /projects/orctest" as shown in the example below, where " /projects/orctest" is a project space)
[UserId@hopper2 ~]$ cd /scratch/UserId
[UserId@hopper2 UserId]$ cp -p -r * /projects/orctest
Optimizing your GPU runs
Current available GPU node options
Type of GPU | Slurm setting | No. of GPUs on Node | No. of CPUs | RAM |
---|---|---|---|---|
A100 80GB | --gres=gpu:A100.80gb:nGPUS | 4 | 64 | 500GB |
DGX A100 40GB | --gres=gpu:A100.40gb:nGPUs | 8 | 128 | 1TB |
The way the GPU nodes are partitioned will likely change over time to optimize utilization.
The best way to take advantage of this Multi-Instance GPU (MIG) mode operation is to analyze the demands of your job and determine which GPU slice is available and suitable for it. For example, if your simulation uses very small GPU memory, you would be better off using a 1g.5gb (where 5GB is the memory you get in the GPU) slice and leaving the bigger slices to jobs that need more GPU memory. Another consideration for machine learning jobs is the difference in demands of training and inference tasks. Training tasks require more computation and memory, therefore they perform best on a full GPU node or a large slice, whereas inference tasks can be adequately performed on smaller slices.
You would modify your Slurm script so that you are now requesting a suitable GPU slice:
#!/bin/bash
#SBATCH --partition=gpuq # the DGX only belongs in the 'gpu' partition
#SBATCH --qos=gpu # need to select 'gpu' QoS
#SBATCH --job-name=python-gpu
## Separate output and error messages into 2 files.
## NOTE: %u=userID, %x=jobName, %N=nodeID, %j=jobID
#SBATCH --output=/scratch/%u/%x-%N-%j.out # Output file
#SBATCH --error=/scratch/%u/%x-%N-%j.err # Error file
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=1 # up to 128;
#SBATCH --gres=gpu:1g.10gb:1 # request a slice of the GPU
#SBATCH --mem-per-cpu=3500M # memory per CORE; total memory is 1 TB (1,000,000 MB)
#SBATCH --export=ALL
#SBATCH --time=0-01:00:00 # set to 1hr; please choose carefully
set echo
umask 0027
# to see ID and state of GPUs assigned
nvidia-smi
module load gnu10
module load python
python myscript.py
Read more about the Hopper GPU nodes and other examples on the DGX USER GUIDE.
Running Parallel Python Jobs
When working on a cluster computer, it is natural to ask how to take
advantage of all these nodes and cores in order to speed up
computation as much as possible. On a laptop, one common approach is to
use the Pool
class in the Python multiprocessing
library to
distribute computation to other cores on the machine. While this
approach certainly works on a cluster too, it does not allow you to take
full advantage of the available computing power. Each job is limited to
a single node and all the cores that are currently available on it.
Multithreaded Python Job
Below is an example Slurm script that can be used to run a Python script that implements the 'multiprocessing' module.
#!/bin/sh
## Give your job a name to distinguish it from other jobs you run.
#SBATCH --job-name=threaded
#SBATCH --partition=normal # will run on any cpus in the 'normal' partition
## Separate output and error messages into 2 files.
## NOTE: %u=userID, %x=jobName, %N=nodeID, %j=jobID
#SBATCH --output=/scratch/%u/%x-%N-%j.out # Output file
#SBATCH --error=/scratch/%u/%x-%N-%j.err # Error file
#SBATCH --constraint=amd ## or intel
#SBATCH --nodes=1 ## all threads need to be on a single node
#SBATCH --cpus-per-task=24 ## 48 or 64 processor
#SBATCH --mem=5G # Total memory needed per task (units: K,M,G,T)
#SBATCH --time=0-01:00:00 # set to 1hr; please choose carefully
set echo
umask 0027
# to see ID and state of GPUs assigned
nvidia-smi
## Load the relevant modules needed for the job
module load gnu10
module load python
## Run your program or script
python <your-threaded-script>.py
The Python script below which can be downloaded here: multithreaded.py can be used as a test case for threaded jobs in Python.
#!/usr/bin/env python
import numpy as np
import multiprocessing as mp
if __name__ == '__main__':
np.random.seed(0);
# create two matrices to be passed
# to two different processes
mat1 = np.random.rand(3,3);
mat2 = np.random.rand(2,2);
# define number of processes
ntasks =2;
# create a pool of processes
p = mp.Pool(ntasks);
# feed different process with same task
# but different data and print the result
print(p.map(np.linalg.eigvals,[mat1,mat2]))
Distributed Python Jobs with mpi4py
The mpi4py
library has a Pool-like class that is very similar to the
one in the multiprocessing
library. Here, we describe how to setup a Python virtual environment
to use mpi4py
run Python code to take advantage
of a much larger number of cores.
Installing mpi4py in a Python Virtual Environment
When installing Python modules, we recommend using a Python Virtual Environment. When working on a project you may want to install a number of different packages. We recommend creating one VE for each project and installing everything that you need into it.
For the purposes of this demonstration, let’s create a virtual
environment called MPIpool and install mpi4py
into it.
[UserId@hopper1 ~]$ module load gnu10
[UserId@hopper1 ~]$ module load python
[UserId@hopper1 ~]$ module load openmpi4/4.1.2
[UserId@hopper1 ~]$ python -m venv ~/MPIpool
[UserId@hopper1 ~]$ source ~/MPIpool/bin/activate
(MPIpool) [UserId@hopper1 ~]$ pip install mpi4py
Collecting mpi4py
Using cached https://files.pythonhosted.org/packages/04/f5/a615603ce4ab7f40b65dba63759455e3da610d9a155d4d4cece1d8fd6706/mpi4py-3.0.2.tar.gz
Installing collected packages: mpi4py
Running setup.py install for mpi4py ... done
Successfully installed mpi4py-3.0.2
Using MPIPoolExecutor in a Python Program
Here we have a sample Python program (which can be downloaded here: MPIpool.py ) that calculates prime numbers. It
uses the MPIPoolExecutor
class to farm out calculations to "workers".
The workers can be running on any node and core in the cluster. There
must always be one "manager" that is responsible for farming out the
work, and collecting the results when finished.
# MPIpool.py
from mpi4py.futures import MPIPoolExecutor
import math
import textwrap
def calc_primes(range_tuple):
"""Calculate all the prime numbers in the given range."""
low, high = range_tuple
if low <= 2 < high:
primes = [2]
else:
primes = []
start = max(3,low) # Don't start below 3
if start % 2 == 0: # Make sure start is odd, i.e. skip evens
start += 1
for num in range(start, high, 2): # increment by 2's, i.e. skip evens
if all(num % i != 0 for i in range(3, int(math.sqrt(num)) + 1, 2)):
primes.append(num)
return primes
def determine_subranges(fullrange, num_subranges):
"""
Break fullrange up into smaller sets of ranges that cover all
the same numbers.
"""
subranges = []
inc = fullrange[1] // num_subranges
for i in range(fullrange[0], fullrange[1], inc):
subranges.append( (i, min(i+inc, fullrange[1])) )
return( subranges )
if __name__ == '__main__':
fullrange = (0, 100000000)
num_subranges = 1000
subranges = determine_subranges(fullrange, num_subranges)
executor = MPIPoolExecutor()
prime_sets = executor.map(calc_primes, subranges)
executor.shutdown()
# flatten the list of lists
primes = [p for plist in prime_sets for p in plist]
print(textwrap.fill(str(primes),80))
The main work is done in the calc_primes()
function, which is what the
workers run. It calculates all the prime numbers within a range defined
by rangeTuple
, a vector that contains two values: the lower and upper
bounds of the range.
The rest of the code runs on the "manager". It calls the
determine_subranges()
function to define the different pieces of work
to send to the workers. The MPIPoolExecutor.map()
function actually
handles all the complexity of coordinating communications with workers,
farming out the different tasks, and then collecting the results.
The mpi4py
documentation suggest that when
using MPIPoolExecutor
, your code should use the
if __name__ == '__main__':
code construct at the bottom of your main
file in order to prevent workers from spawning more
workers.
Submitting the Program to Slurm
Here we provide a Slurm script for running such a job.
#!/bin/sh
## Give your job a name to distinguish it from other jobs you run.
#SBATCH --job-name=MPIpool
#SBATCH --partition=normal
## Separate output and error messages into 2 files.
## NOTE: %u=userID, %x=jobName, %N=nodeID, %j=jobID, %A=arrayID, %a=arrayTaskID
#SBATCH --output=/scratch/%u/%x-%N-%j.out # Output file
#SBATCH --error=/scratch/%u/%x-%N-%j.err # Error file
## Slurm can send you updates via email
#SBATCH --mail-type=BEGIN,END,FAIL # ALL,NONE,BEGIN,END,FAIL,REQUEUE,..
#SBATCH --mail-user=<GMUnetID>@gmu.edu # Put your GMU email address here
## Specify how much memory your job needs. (2G is the default)
#SBATCH --mem=8G # Total memory needed per task (units: K,M,G,T)
## Specify how much time your job needs. (default: see partition above)
#SBATCH --time=0-02:00 # Total time needed for job: Days-Hours:Minutes
#SBATCH --ntasks=51 # 50 workers, 1 manager
set echo
umask 0027
# to see ID and state of GPUs assigned
nvidia-smi
## Load the relevant modules needed for the job
module load gnu10
module load python
module load openmpi4/4.1.2
source ~/MPIpool/bin/activate
## Run your program or script
mpirun -np $SLURM_NTASKS python -m mpi4py.futures MPIpool.py
Be sure to replace the <
and >
) with you own email address.
Note that we set --ntasks=51
in order to allocate 1 manager and 50
workers. There must always be only 1 manager and at least 1 worker. Note
that we use the $SLURM_NTASKS
environment variable in the call to
mpirun
to make sure that the number of cores used equals the number
allocated by the --ntasks=
option.
Because mpi4py
is based on the MPI libraries, we need to load one of
the MPI modules. Here we have chosen OpenMPI. The mpirun
or mpiexec
program must be used to properly launch an MPI program, and this program
is no exception.
The runtime for this program using 50 workers is about 1 minute. That is significantly faster than the 45 minutes needed to run the program using a single core. Of course, there is a point of diminishing returns (and even an added cost) in adding more and more workers. It is good to experiment with different numbers to see how many workers are optimal. The maximum number of cores that a user can request is currently 300. This may change in the future.
This is an example of an algorithm that is "embarrassingly parallel". It is very easy to divide it up into smaller pieces and pass them out. Many algorithms are not so easy to parallelize in this way. MPI is a very mature library, and it has the tools to handle problems that are much more complex than this. It is the de facto standard for doing large scale parallelization, and if that is your goal you can benefit from learning more about it. Those interested in a more "Pythonic" library may want to look into Dask.
Using External Python Packages
To install and run your Python code with external Python packages, after loading the Python module, first
create a directory for storing those packages (e.g. ~/python-packages/projectX
)
mkdir ~/python-packages
mkdir ~/python-packages/projectX
Then install the appropriate packages in there:
pip install <package1> -t ~/python-packages/projectX
To run your code with these extra packages, you would need to add the export
command to your
Slurm submission script so that the last part would now be
module load gnu10
module load python
export PYTHONPATH=~/python-packages/projectX:$PYTHONPATH
python myscript.py
export
command from the terminal
$ export PYTHONPATH=~/python-packages/projectX:$PYTHONPATH
Running with Python Virtual Environments
To have better control over the Python packages and libraries you need to use on the cluster, the best way to run Python is using Python Virtual Environments. This is especially useful for codes that use Tensorflow, Keras or Pytorch. Read our instructions on building Python Virtual Environments and how to run Tensorflow.
Remember
When running on Argo, the Slurm scripts have to be updated so that they can be run on Argo. The main differences between Argo and Hopper are detailed in these pages.
Running with Jupyter NoteBooks
You also have the option of using Jupyter Notebooks (on Hopper) to run Python code. The steps for doing this are outlined in these pages.