Theta (Argonne LCF)¶
Theta is a 11.69 petaflops system based on the second-generation Intel Xeon Phi processor at Argonne Leadership Computing Facility (ALCF). It serves as a stepping stone to the ALCF’s next leadership-class supercomputer, Aurora. Theta is a massively parallel, many-core system based on Intel processors and interconnect technology, a new memory space, and a Lustre-based parallel file system, all integrated by Cray’s HPC software stack.
Already installed module¶
This installation procedure shows you how to access the installed DeepHyper module on Theta. After logging in Theta, to access Deephyper run the following commands:
$ module load conda/2021-09-22 $ conda activate base
Then to verify the installation do:
$ python >>> import deephyper >>> deephyper.__version__ '0.3.0'
This installation procedure shows you how to create your own Conda virtual environment and install DeepHyper in it.
It is important to run the following commands from the appropriate storage space because some features of DeepHyper can generate a consequent quantity of data such as model checkpointing. The storage spaces available at the ALCF are:
For more details refer to ALCF Documentation.
After logging in Theta, go to your project folder (replace
PROJECTNAME by your own project name):
$ cd /lus/theta-fs0/projects/PROJECTNAME
Then create the
$ module load miniconda-3 $ conda create -p dhknl python=3.8 -y $ conda activate dhknl/
It is then required to have the following additionnal dependencies:
$ conda install gxx_linux-64 gcc_linux-64 -y
Finally install DeepHyper in the previously created
$ pip install pip --upgrade $ # DeepHyper + Analytics Tools (Parsing logs, Plots, Notebooks) $ pip install deephyper[analytics] $ conda install tensorflow -c intel -y
Horovod can be installed to use data-parallelism during the evaluations of DeepHyper. To do so use
pip install deephyper[analytics,hvd] while or after installing.
To use Jupyter notebooks on Theta go to Theta Jupyter and use your regular authentication method. The Jupyter Hub tutorial from Argonne Leadership Computing Facility might help you in case of troubles.
To create a custom Jupyter kernel run the following from your activated Conda environment:
$ python -m ipykernel install --user --name deephyper --display-name "Python (deephyper)"
Now when openning a notebook from Jupyter Hub at ALCF make sure to use the
Python (deephyper) kernel before executing otherwise you will not have all required dependencies.