brc-slurm

This is the manual for the head start at berkeley research computing cluster

logging in

ssh [email protected]

working place

• /global/home/username only have 10 GB limit
• /global/scratch/<username> has infinite storage

transferring data

# sending to the server
scp (-r) local_path/A [email protected]:path/A
# receive from the server
scp (-r) [email protected]:path/A local_path/A 

Making vim available!

wget https://raw.githubusercontent.com/amix/vimrc/master/vimrcs/basic.vim
mv basic.vim ~/.vimrc

vim basics

• i: insert before the cursor
• Esc: exit insert mode
• Esc+:w: write (save) the file, but don't exit
• Esc+:q: quit (fails if there are unsaved changes)
• Esc+:q!: quit and throw away unsaved changes

Install the anaconda (OUTDATED!)

• Download the anaconda

wget https://repo.anaconda.com/archive/Anaconda3-2020.11-Linux-x86_64.sh
bash Anaconda3-2020.11-Linux-x86_64.sh

• add the path

echo 'export PATH=“/global/scratch/<username>/<anaconda-path>/bin:$PATH”' >> ~/.bashrc
source ~/.bashrc

use anaconda (NEW~)

module load python

Then, create the conda environment as usual.

Since the disk quota in the local storage is limited, we need to change the location using the following commands (https://stackoverflow.com/questions/67610133/how-to-move-conda-from-one-folder-to-another-at-the-moment-of-creating-the-envi)

# create a new pkgs_dirs (wherever, doesn't have to be hidden)
mkdir -p /big_partition/users/user/.conda/pkgs

# add it to Conda as your default
conda config --add pkgs_dirs /big_partition/users/user/.conda/pkgs

# create a new envs_dirs (again wherever)
mkdir -p /big_partition/users/user/.conda/envs

# add it to Conda as your default
conda config --add envs_dirs /big_partition/users/user/.conda/envs

• use cuda on brc

module load cuda/10.2
export XLA_FLAGS=--xla_gpu_cuda_data_dir=/global/software/sl-7.x86_64/modules/langs/cuda/10.2

adding the software

• module avail - List all available modulefiles.
• module list - List modules loaded.
• module add|load _modulefile_ ... - Load modulefile(s).
• module rm|unload _modulefile_ ... - Remove modulefile(s).

Example

if the code uses matlab, make sure you load the matlab module: module load matlab

Running the jobs

• sbatch myjob.sh - Submit a job, where myjob.sh is a SLURM job script.
• squeue -u $USER - Check your current jobs
• scancel [jobid] - Cancel a job with a job ID from `squeue -u $USER`
• sinfo - View the status of the cluster's compute nodes

• sacctmgr -p show associations user=$USER --- show which partition can be used in the account

One example of myjob.sh

#!/bin/bash
# Job name:
#SBATCH --job-name=test
#
# Account:
#SBATCH --account=co_esmath
#
# Partition:
#SBATCH --partition=savio3
#
# Quality of Service:
#SBATCH --qos=esmath_savio3_normal
# Number of nodes: 
#SBATCH --nodes=1
# Processors per task 
#SBATCH --cpus-per-task=2
#
# Wall clock limit:
#SBATCH --time=24:00:00
# Email Notification
#SBATCH --mail-type=END, FAIL
#SBATCH [email protected]
#
## Command(s) to run:


# load some necessary software
module load matlab mpi 

# if one use conda for the python environment
conda activate myenv

# run my jobs
bash myscript.sh

# python jobs
python myscript.py

# matlab jobs
matlab < main.m 

One example of myjob.sh (GPU Instance)

#!/bin/bash
# Job name:
#SBATCH --job-name=test
#
# Account:
#SBATCH --account=co_esmath
#
# Partition:
#SBATCH --partition=savio3_gpu
#
# Quality of Service:
#SBATCH --qos=esmath_gpu3_normal
# Number of nodes: 
#SBATCH --nodes=1
# Processors per task 
#SBATCH --cpus-per-task=2
#
#SBATCH --gres=gpu:GTX2080TI:1
# Wall clock limit:
#SBATCH --time=24:00:00
# Email Notification
#SBATCH --mail-type=END, FAIL
#SBATCH [email protected]
#
## Command(s) to run:

# load gpu related 
module load gcc openmpi
module load cuda/11.2
module load cudnn/7.0.5
export CUDA_PATH=/global/software/sl-7.x86_64/modules/langs/cuda/11.2
export LD_LIBRARY_PATH=$CUDA_PATH/lib64:$LD_LIBRARY_PATH
 
# if one use conda for the python environment
conda activate myenv

# python jobs
XLA_FLAGS=--xla_gpu_cuda_data_dir=/global/software/sl-7.x86_64/modules/langs/cuda/11.2 python myscript.py

• you can find the hardware config

advanced usage

pip install sysflow

config the slurm

slurm config 

run the jobs

slurm run [python test.py --arg1 5 --arg2 3]

examples in python

from sysflow.job.slurm import Slurm

# use the last config
slurm = Slurm()

# change the config 
# slurm = Slurm(job_name='hello-word', email='[email protected]', conda_env='qrl')

# change the account or partition
# slurm = Slurm(account='co_esmath', qos='esmath_savio3_normal', partition='savio3')

slurm.run('python test.py')

slurm config --account fc_esmath --qos savio_normal
slurm config --account co_esmath --qos esmath_savio3_normal --partition savio3 --task_per_node 32
slurm config --account co_esmath --qos savio_lowprio --partition savio2 --task_per_node 24

Reference

• official reference
• others (chinese)
• cheatsheet