This guide compares performance of tSNE alogrithm on CPU platform vs GPU. For GPU, we are using Nvidia RAPIDSAI Framework. In this example, rapidsai container was used with Singularity.
Various datasets were used for this study. Sources of datasets along with their usage guidelines is listed under 'datasets' directory
# On your local machine, establish ssh connection to raad2 GPU cluster with Port Forwarding.
ssh -L 8000:localhost:8000 muarif092@raad2-gfx.qatar.tamu.edu
# Once you are logged into supercomputer, submit an interactive job to get hold of a GPU node
srun --ntasks=18 --gres=gpu:v100:1 --time=08:00:00 --tunnel L8000:8000 --pty /bin/bash
# Now load CUDA runtime
module load cuda10.0
# Clone the test code (needs to be done once)
git clone https://github.com/mustafaarif/tSNE.git
cd tSNE/
mkdir data
# Download data from above mentioned source and place in data directory.
# This data contains few missing values. These values can be replaced with 0.
# Start Container. We are using Nvidia provided RAPIDS container
singularity shell /cm/shared/apps/rapidsai/rapids0.13-cuda10.0-centos7
# Activate RAPIDS conda environment
source activate rapids
# Launch Jupyter notebook
jupyter lab --port 8000For breastCancer dataset, tSNE fit for CPU version took 2.89sec whereas GPU version took 0.89sec.