1. Login to rishon
   ssh -X example@rishon.cs.technion.ac.il
2. Print partitions(queues states)
   sinfo
3. Print jobs states
   squeue -o “%.18i %.9P %.8j %.8u %.2t %.10M %.6D %R %Q”
4. Ask for bash shell in one of the computing servers
   srun -p 236605,all -c2 --gres=gpu:1 --pty bash
   Explanation:
   -p 236605,all  submit the job to two queues(all,236605) , once it gets elected to run in one of them , it automatically removed in the other. this way the user benefit using his higher priorty on his course queue(236605 or 236370) and if this queue is too busy it can ran with lower priority on other servers in all queue.
   make sure you write the course number you enroll to.People who are not enroll to either of these courses, should not run with this flag, the system will route them according to their status.c2 – ask for 2 cores
   gpu:1 – ask for 1 gpu of any type
   --pty bash – gives terminal in one of the computing node.
5. See your job in the queue
   squeue -o “%.18i %.9P %.8j %.8u %.2t %.10M %.6D %R %Q”
6. Edit .bashrc file
   enter ~/.bashrc file with:
   nano ~/.bashrc
   add as last line the following:
   export LC_ALL=”en_US.UTF-8″
7. Install development environment (notice the --user ! )
   This is an example for tensorflow + pytorch + keras, change it for your own needs.
   Only run in computing node, not on login server(rishon)
   %pip3 for python 3.5 or pip2 for python 2.7
pip3 install --user --upgrade pip
pip3 install --user --upgrade tensorflow-gpu==1.11
pip3 install --user http://download.pytorch.org/whl/cu90/torch-0.3.1-cp35-cp35m-linux_x86_64.whl
(pip2 install --user http://download.pytorch.org/whl/cu90/torch-0.3.1-cp27-cp27mu-linux_x86_64.whl)
pip3 install --user torchvision
pip3 install --user numpy scipy matplotlib keras

* if pip3/pip2 doesn’t work , use: python3/2 -m pip install –user ….

• Login to rishon
  ssh -X example@rishon.cs.technion.ac.il
• Create exp2.bash, using nano or vi
  #!/bin/bash
  #SBATCH -N 1  # number of  minimum nodes
  #SBATCH -c 2 # number of cores
  #SBATCH –gres=gpu:1   # Request 1 gpu
  #SBATCH -p 236370,all #  or 236605,all for more details see explanation above (srun command)
  #SBATCH –mail-user=example@cs.technion.ac.il
  # (change to your own email if you wish to get one, or just delete this and the following lines)
  #SBATCH –mail-type=ALL  # Valid values are NONE, BEGIN, END, FAIL, REQUEUE, ALL
  #SBATCH –job-name=”just_a_test”
  #SBATCH -o slurm.%N.%j.out    # stdout goes here
  #SBATCH -e slurm.%N.%j.out   # stderr goes here
  nvidia-smi
  source ~/tensorflow/bin/activate
  python3 main.py
• From rishon server run
  sbatch exp2.bash
• Lookit up in queue list
  squeue -o “%.18i %.9P %.8j %.8u %.2t %.10M %.6D %R %Q”
• View it output
  tail -f ~/slurm.NodeName.JobID.out (fill in your NodeName,JobID)
  Example output:
  tail -f slurm.rishon3.103512.out
  Train Epoch: 3 [53760/60000 (90%)]      Loss: 0.172956
  Train Epoch: 3 [54400/60000 (91%)]      Loss: 0.340682
  Train Epoch: 3 [55040/60000 (92%)]      Loss: 0.294907
  Train Epoch: 3 [55680/60000 (93%)]      Loss: 0.243409
  Train Epoch: 3 [56320/60000 (94%)]      Loss: 0.262233
  Train Epoch: 3 [56960/60000 (95%)]      Loss: 0.220493
  Train Epoch: 3 [57600/60000 (96%)]      Loss: 0.375372
  Train Epoch: 3 [58240/60000 (97%)]      Loss: 0.156547
  Train Epoch: 3 [58880/60000 (98%)]      Loss: 0.272710
  Train Epoch: 3 [59520/60000 (99%)]      Loss: 0.129080

Troubleshoot: If the output seems wrong, you can see the output live using srun instead of sbatch.

srun -c2 -N1 --gres=gpu:1 --job-name="just_a_test" exp2.bash

Don’t forget to chmod +x exp2.bash.

עבודה עם ג’ופיטר

התקנה:

python3 -m pip install --upgrade --user pip
python3 -m pip install --user jupyterlab


השלב הבא הוא ליצור סקריפט שיחסוך לנו הקלדות, אנא צרו קובץ בשם jl.bash עם התוכן הבא:

#!/bin/bash
hn=`hostname -I`
unset XDG_RUNTIME_DIR
.local/bin/jupyter-lab --no-browser --ip=$hn --port=PORT


נא לבחור PORT כמספר בתחום 2000 ומעלה

example@gaon:~$chmod +x ./jl.bash

ועכשיו אפשר לקבל ג’ופיטר לאב על המחשב שלנו שרץ על השרתי חישוב

example@gaon:~$srun -c4 --gres=gpu:1 --pty jl.bash


את הכתובת שמקבלים יש כמובן להעתיק לדפדפן המקומי על המחשב שלנו