Jupyter Notebooks for HPC
tbd: Jens Henrik Goebbert @goebbert1
To access the JSC JupyterLab, the link is: http://jupyter-jsc.fz-juelich.de
It is important to know what and where you want to run. For ML and DL workloads, it is advisable to use the GPU nodes of the supercomputers. When choosing the parameters for a run, it is important to use the "gpus" or "develgpus" partitions. If you are not using parallel environments (such as Horovod), choose only one node of the Develgpus partition: You will get an allocation faster, and other people can also work on it.
Always choose the maximum number of GPUs allowed.
Given that this is a shared environment, it might take a while until you get a reservation. Therefore, there is a new option to receive an email when your reservation is granted.
Setup of common tools for HPC
TensorFlow and Horovod for ImageNet
tbd: Jenia Jitsev @jitsev1
JURECA and JUWELS
Fahad Khalid (@khalid1): The following Deep Learning related modules are available in the production stage:
1. Tensorflow 1.13.1 2. Keras 2.2.4 3. Horovod 0.16.2
Fahad Khalid (@khalid1): The following Deep Learning related modules are available on JURON:
1. Tensorflow 1.12.0 (Python 2 and Python 3) 2. Keras 2.2.4 (Python 2 and Python 3) 3. PyTorch 1.0.1 (Python 3) 4. Horovod 0.15.2 (Python 2 and Python 3) 5. Caffe 1.0 (Python 2 and Python 3)
All thanks to Andreas Herten (@herten1) for installing these modules and the many dependencies.
Pytorch & HEAT
tbd: Bjoern Hagemeier @hagemeier2