Last updated July 2021.

First, login to Traverse cluster headnode via ssh:

ssh -XC <yourusername>@traverse.princeton.edu

Note, -XC is optional; it is only necessary if you are planning on performing remote visualization, e.g. the output .png files from the below section. Trusted X11 forwarding can be used with -Y instead of -X and may prevent timeouts, but it disables X11 SECURITY extension controls. Compression -C reduces the bandwidth usage and may be useful on slow connections.

Add the following flag to your environment.

export LD_PRELOAD=/usr/lib64/libpmix.so.2

The recommended way is to edit your ~/.bashrc and then reload the environment as follows:

source ~/.bashrc

Next, check out the source code from github:

git clone https://github.com/PPPLDeepLearning/plasma-python
cd plasma-python
git checkout tf2

After that, create an isolated Anaconda environment and load CUDA drivers, an MPI compiler, and the HDF5 library. The suggested libraries are included with the repository, so simple sourcing plasma-python/envs/traverse.cmd will load these into your path. Traverse uses a PowerPC arcitecture, which is not as widely supported with many common libraries. The python=3.6 option below sets the python version in this anaconda environment to the older python version 3.6, which will help down the line with installing libraries.

#cd plasma-python
module load anaconda3
conda env create --name my_env --file envs/requirements-traverse.yaml python=3.6
conda activate my_env

Go into envs/traverse.cmd and modify the conda activate command to conda activate my_env. Run:

source envs/traverse.cmd

As of the latest update of this document (Summer 2021), the above modules correspond to the following versions on the Traverse system, given by module list:

Currently Loaded Modulefiles:
  1) anaconda3/2020.7                 3) cudnn/cuda-9.2/7.6.3             5) hdf5/gcc/openmpi-1.10.2/1.10.0
  2) cudatoolkit/10.2                 4) openmpi/cuda-11.0/gcc/4.0.4/64

Next, install the plasma-python package:

#conda activate my_env
python setup.py install --user

When queueing jobs on Traverse or running slurm managed scripts, DO NOT load your anaconda environment before doing so. This will cause a module loading issue. It is highly suggested that you build plasma-python in one terminal witht the anaconda environment loaded and run it in another without the anaconda environment loaded to avoid this issue. Alternatively, calling module purge before using slurm fixes this issue.

On Traverse, pytorch has been observed to not install correctly. By default it is commented out, but if that's not the case the quick fix is to not intall it by commenting out the line 17 in envs/requirements-traverse.yaml

  7 dependencies:
  8   - python>=3.6.8
  9   - cython
 10   - pip
 11   - scipy
 12   - pandas
 13   - flake8
 14   - h5py<3.0.0
 15   - pyparsing
 16   - pyyaml
 17   #- pytorch>1.3

On Traverse, xgboost doesn't seem to build right. Just ignore it for now by editing setup.py (comment line 35). Rebuild as above.

 30       install_requires=[
 31           'pathos',
 32           'matplotlib',
 33           'hyperopt',
 34           'mpi4py',
 35           #'xgboost',
 36           'scikit-learn',
 37           'joblib',
 38           ],

Common issue is Intel compiler mismatch in the PATH and what you use in the module. With the modules loaded as above, you should see something like this (as of summer 2021):

$ which mpicc
/usr/local/openmpi/cuda-11.0/4.0.4/gcc/x86_64/bin/mpicc

In both cases, especially note the presence of the CUDA directory in this path. This indicates that the loaded OpenMPI library is CUDA-aware.

If you conda activate the Anaconda environment after loading the OpenMPI library, your application would be built with the MPI library from Anaconda, which has worse performance on this cluster and could lead to errors. See mpi4py on HPC Clusters for a related discussion.

Tigress is also avilable on Traverse, so this step is identical to TigerGPU. The JET and D3D datasets contain multi-modal time series of sensory measurements leading up to deleterious events called plasma disruptions. The datasets are located in the /tigress/FRNN project directory of the GPFS filesystem on Princeton University clusters.

For convenience, create following symbolic links:

cd /tigress/<netid>
ln -s /tigress/FRNN/shot_lists shot_lists
ln -s /tigress/FRNN/signal_data signal_data

All the configuration parameters are summarised in examples/conf.yaml. In this section, we highlight the important ones used to control the input data.

Currently, FRNN is capable of working with JET and D3D data as well as thecross-machine regime. The switch is done in the configuration file:

paths:
    ... 
    data: 'jet_0D'

Older yaml files kept for archival purposes will denote this data set as follow:

paths:
    ... 
    data: 'jet_data_0D'

use d3d_data for D3D signals, use jet_to_d3d_data ir d3d_to_jet_data for cross-machine regime.

By default, FRNN will select, preprocess, and normalize all valid signals available in the above dataset. To chose only specific signals use:

paths:
    ... 
    specific_signals: [q95,ip] 

if left empty [] will use all valid signals defined on a machine. Only set this variable if you need a custom set of signals.

Other parameters configured in the conf.yaml include batch size, learning rate, neural network topology and special conditions foir hyperparameter sweeps.

On Traverse, the data is stored in the tigress filesystem. You will probably need to modify conf.yaml to point there by setting:

fs_path: '/tigress/'
...
fs_path_output: '/tigress/'

cd examples/
python guarantee_preprocessed.py

This will preprocess the data and save rescaled copies of the signals in /tigress/<netid>/processed_shots, /tigress/<netid>/processed_shotlists and /tigress/<netid>/normalization

Preprocessing must be performed only once per each dataset. For example, consider the following dataset specified in the config file examples/conf.yaml:

paths:
    data: jet_0D

Preprocessing this dataset takes about 20 minutes to preprocess in parallel and can normally be done on the cluster headnode.

Use the Slurm job scheduler to perform batch or interactive analysis on TigerGPU cluster.

For non-interactive batch analysis, make sure to allocate exactly 1 MPI process per GPU. Save the following to slurm.cmd file (or make changes to the existing examples/slurm.cmd):

#!/bin/bash
#SBATCH -t 01:30:00
#SBATCH -N X
#SBATCH --ntasks-per-node=4
#SBATCH --ntasks-per-socket=2
#SBATCH --gres=gpu:4
#SBATCH -c 4
#SBATCH --mem-per-cpu=0

module load anaconda3
conda activate my_env
export OMPI_MCA_btl="tcp,self,vader"
module load cudatoolkit cudann 
module load openmpi/cuda-8.0/intel-17.0/3.0.0/64
module load intel
module load hdf5/intel-17.0/intel-mpi/1.10.0

srun python mpi_learn.py

where X is the number of nodes for distibuted training and the total number of GPUs is X * 4. This configuration guarantees 1 MPI process per GPU, regardless of the value of X.

Update the num_gpus value in conf.yaml to correspond to the total number of GPUs specified for your Slurm allocation.

Submit the job with (assuming you are still in the examples/ subdirectory):

#cd examples
sbatch slurm.cmd

And monitor it's completion via:

squeue -u <netid>

Optionally, add an email notification option in the Slurm configuration about the job completion:

#SBATCH --mail-user=<netid>@princeton.edu
#SBATCH --mail-type=ALL

Interactive option is preferred for debugging or running in the notebook, for all other case batch is preferred. The workflow is to request an interactive session:

salloc -N [X] --ntasks-per-node=4 --ntasks-per-socket=2 --gres=gpu:4 -c 4 --mem-per-cpu=0 -t 0-6:00

Re-load the above modules and reactivate your conda environment. Confirm that the correct CUDA-aware OpenMPI library is in your interactive Slurm sessions's shell search path:

$ which mpirun 
/usr/local/openmpi/cuda-8.0/3.0.0/intel170/x86_64/bin/mpirun

Then, launch the application from the command line:

mpirun -N 4 python mpi_learn.py

where -N is a synonym for -npernode in OpenMPI. Do not use srun to launch the job inside an interactive session. If you an encounter an error such as "unrecognized argument N", it is likely that your modules are incorrect and point to an Intel MPI distribution instead of CUDA-aware OpenMPI. Intel MPI is based on MPICH, which does not offer the -npernode option. You can confirm this by checking:

$ which mpirun 
/opt/intel/compilers_and_libraries_2019.3.199/linux/mpi/intel64/bin/mpirun

A regular FRNN run will produce several outputs and callbacks.

Currently supports graph visualization, histograms of weights, activations and biases, and scalar variable summaries of losses and accuracies.

The summaries are written in real time to /tigress/<netid>/Graph. For macOS, you can set up the sshfs mount of the /tigress filesystem and view those summaries in your browser.

To install SSHFS on a macOS system, you could follow the instructions here: https://github.com/osxfuse/osxfuse/wiki/SSHFS Or use Homebrew, brew cask install osxfuse; brew install sshfs. Note, to install and/or use osxfuse you may need to enable its kernel extension in: System Preferences → Security & Privacy → General

After installation, execute:

sshfs -o allow_other,defer_permissions [email protected]:/tigress/<netid>/ <destination folder name on your laptop>/

The local destination folder may be an existing (possibly nonempty) folder. If it does not exist, SSHFS will create the folder. You can confirm that the operation succeeded via the mount command, which prints the list of currently mounted filesystems if no arguments are given.

Launch TensorBoard locally (assuming that it is installed on your local computer):

python -m tensorboard.main --logdir <destination folder name on your laptop>/Graph

A URL should be emitted to the console output. Navigate to this link in your browser. If the TensorBoard interface does not open, try directing your browser to localhost:6006.

You should see something like:

When you are finished with analyzing the summaries in TensorBoard, you may wish to unmount the remote filesystem:

umount  <destination folder name on your laptop>

The local destination folder will remain present, but it will no longer contain the remote files. It will be returned to its previous state, either empty or containing the original local files. Note, the umount command is appropriate for macOS systems; some Linux systems instead offer the fusermount command.

These commands may be useful when the SSH connection is lost and an existing mount point cannot be re-mounted, e.g. errors such as:

mount_osxfuse: mount point <destination folder name on your laptop> is itself on a OSXFUSE volume

More aggressive options such as umount -f <destination folder name on your laptop> and alternative approaches may be necessary; see discussion here.

Besides TensorBoard summaries, you can visualize the accuracy of the trained FRNN model using the custom Python scripts and notebooks included in the repository.

You can produce the ROC curves for validation and test data as well as visualizations of shots by using:

cd examples/
python performance_analysis.py

The performance_analysis.py script uses the file produced as a result of training the neural network as an input, and produces several .png files with plots as an output.

In addition, you can check the scalar variable summaries for training loss, validation loss, and validation ROC logged at /tigress/<netid>/csv_logs (each run will produce a new log file with a timestamp in name).

Sample notebooks for analyzing the files in this directory can be found in examples/notebooks/. For instance, the LearningCurves.ipynb notebook contains a variation on the following code snippet:

import pandas as pd
import numpy as np
from bokeh.plotting import figure, show, output_file, save

data = pd.read_csv("<destination folder name on your laptop>/csv_logs/<name of the log file>.csv")

from bokeh.io import output_notebook
output_notebook()

from bokeh.models import Range1d
#optionally set the plotting range
#left, right, bottom, top = -0.1, 31, 0.005, 1.51

p = figure(title="Learning curve", y_axis_label="Training loss", x_axis_label='Epoch number') #,y_axis_type="log")
#p.set(x_range=Range1d(left, right), y_range=Range1d(bottom, top))

p.line(data['epoch'].values, data['train_loss'].values, legend="Test description",
       line_color="tomato", line_dash="dotdash", line_width=2)
p.legend.location = "top_right"
show(p, notebook_handle=True)

The resulting plot should match the train_loss plot in the Scalars tab of the TensorBoard summary.

To extract per mini-batch summaries, we require a finer granularity of checkpoint data than what it is logged to the per-epoch lines of csv_logs/ files. We must directly use the output produced by FRNN logged to the standard output stream. In the case of the non-interactive Slurm batch jobs, it will all be contained in the Slurm output file, e.g. slurm-3842170.out. Refer to the following notebook to perform the analysis of learning curve on a mini-batch level: FRNN_scaling.ipynb