如何将SLURM作为输入参数传递给python？

Question

我是用SLURM训练一批卷积神经网络的新手。为了跟踪所有训练好的CNN，我想把SLURM jobID作为输入参数传递给python。将其他变量作为参数传递很好。但是，我无法访问SLURM作业人员的身份。

我已经试过使用${SLURM_JOBID}、${SLURM_JOB_ID}、%j和%J了。在传递到python之前，我还尝试将这些slurm变量写入变量中。

以下是我的最新代码：

#!/bin/bash

# --- info to user
echo "script started ... "

# --- setup environment
module purge            # clean up
module load python/3.6
module load nvidia/10.0
module load cudnn/10.0-v7 

# --- display information
HOST=`hostname`
echo "This script runs the CNN. Slurm scheduled it on node $HOST"
echo "I am interested of all environment variables Slurm adds:"
env | grep -i slurm

# --- start running ... 
echo " --- run --- "

# --- define some varibles
dc="dice"
sm="softmax"

# --- run a job using a slurm batch script
for layer in {3..15..2}
  do
    sbatch -N 1 -n 1 --mem=20G --mail-type=END --gres=gpu:V100:3 --wrap="singularity --noslurm tensorflow_19.03-py3.simg python run_CNN_dynlayer.py ${SLURM_JOBID} ${layer} ${dc}"
    sleep 1 # pause 1s to be kind to the scheduler...
    echo "jobid: "+${SLURM_JOBID}
    echo " --- next --- "
  done    

cmd看起来是这样的：

femonk@rarp1 [CNN] ./run_CNN_test.slurm
script started ... 
This script runs the CNN. Slurm scheduled it on node rarp1
I am interested of all environment variables Slurm adds:
SLURM_ACCOUNT=AI
PYTHONPATH=/cluster/slurm/lib64/python3.6/site-packages:/cluster/slurm/lib64/python3.6/site-packages:/cluster/slurm/lib64/python3.6/site-packages:
 --- run --- 
Submitted batch job 3182711
jobid: 
 --- next --- 
femonk@rarp1 [CNN] 

有人知道我的代码出了什么问题吗？提前谢谢。

Answer 1

SLURM_JOBID环境变量仅用于作业流程，而不是提交作业的流程。作业id是从sbatch命令返回的，因此如果您希望它在变量中，则需要分配它。

  do
    SLURM_JOBID=$(sbatch --parsable -N 1 -n 1 --mem=20G --mail-type=END --gres=gpu:V100:3 --wrap="singularity --noslurm tensorflow_19.03-py3.simg python run_CNN_dynlayer.py ${SLURM_JOBID} ${layer} ${dc}")
    sleep 1 # pause 1s to be kind to the scheduler...
    echo "jobid: "+${SLURM_JOBID}
    echo " --- next --- "
  done   

注意命令替换$()与sbatch的--parsable参数的结合使用。

还请注意，当前输出的行Submitted batch job 3182711将消失，因为它用于填充SLURM_JOBID变量。