Using Latte on Cori

This guide walks through setting up Latte on the Cori supercomputer.

Quick links

• Module Dependencies
• Building Latte
• CIFAR-10 Example
• Tiny-imagenet Example
• ImageNet Example

Module Dependencies

module load julia hdf5-parallel impi

Add these to your ~/.bash_profile.ext to load automatically on login.

Building Latte

$ julia -e 'Pkg.checkout("CompilerTools")'
$ julia -e 'Pkg.checkout("ParallelAccelerator")'
$ julia -e 'Pkg.clone("https://github.com/IntelLabs/Latte.jl")'

Build Latte's dependencies with MPI enabled

# Build the IO library
cd ~/.julia/v0.4/Latte/deps
cd IO
cmake -DLATTE_MPI=ON . && make
cp libLatteIO.so ../

# Build the communication library
cd ../communication
cmake . && make
cp libLatteComm.so ../

CIFAR-10 Example

Download and convert the dataset into hdf5

cd ~/.julia/v0.4/Latte/examples/cifar10/data
./get_data.sh $SCRATCH

Create a SLURM batch script called train_vgg_mini.sh

#!/bin/bash -l
#SBATCH -p debug
#SBATCH -N 2
#SBATCH -t 02:00:00
#SBATCH -J train_vgg_mini_cifar
srun -n 2 julia vgg-mini.jl

Submit a job

sbatch train_vgg_mini.sh

Tiny-imagenet Example

Build the image conversion utility

cd ~/.julia/v0.4/Latte/utils/converter
cmake . && make
cp convert_imagenet ../../bin

Download and convert the dataset

cd ~/.julia/v0.4/Latte/examples/tiny-imagenet/data
./get-data.sh $SCRATCH

ImageNet Example

Visit the ImageNet website to download the ILSVRC2012 training and validation images for Task 1 to $SCRATCH.