Prisma

Prisma is a storage data plane that accelerates the training performance of Deep Learning (DL) frameworks. It implements a parallel data prefetching mechanism that reads training data in advance and stores it in an in-memory buffer to serve incoming I/O requests of DL frameworks.

Please cite the following paper if you use Prisma:

The Case for Storage Optimization Decoupling in Deep Learning Frameworks. Ricardo Macedo, Cláudia Correia, Marco Dantas, Cláudia Brito, Weijia Xu, Yusuke Tanimura, Jason Haga, João Paulo. IEEE International Conference in Cluster Computing @ 1st Workshop on Re-envisioning Extreme-Scale I/O for Emerging Hybrid HPC Workloads.

@inproceedings {Macedo2021Prisma,
  title     = {{The Case for Storage Optimization Decoupling in Deep Learning Frameworks}},
  author    = {Ricardo Macedo and Cl{\'{a}}udia Correia and Marco Dantas and Cl{\'{a}}udia Brito and Weijia Xu and Yusuke Tanimura and Jason Haga and Jo{\~{a}}o Paulo},
  booktitle = {{IEEE International Conference on Cluster Computing}},
  pages     = {649--656},
  year      = {2021},
  doi       = {10.1109/Cluster48925.2021.00096},
  publisher = {{IEEE}}
}

Getting started with Prisma

This tutorial will guide you through the installation and configuration of Prisma.

Install dependencies

Prisma was built, compiled, and tested with g++-9.3.0 and cmake-3.16. It has dependencies with Boost C++ libraries, so please install them before running the build commands. Prisma also has dependencies with Intel Threading Building Blocks (TBB), but these are installed (dynamically) at compile time.

$ wget https://boostorg.jfrog.io/artifactory/main/release/1.76.0/source/boost_1_76_0.tar.gz
$ tar xzvf boost_1_76_0.tar.gz
$ cd boost_1_76_0/
$ ./bootstrap.sh --prefix=/opt 
$ ./b2
$ sudo ./b2 install
$ grep "#define BOOST_LIB_VERSION" /opt/include/boost/version.hpp

Install Prisma

To build Prisma, run the following commands:

$ git clone https://github.com/dsrhaslab/prisma.git
$ cd prisma/prisma/build
$ cmake ..
$ make

After running the commands, the shared library libprisma.so will appear in the build directory.

Create the filenames list

Before executing Prisma, a filenames_list must be created, so that Prisma knows the files that must be read in advance, and the order in which they must be cached.

To create the filenames_list, use the shuffle_filenames/shuffle_filenames.py script, by running the following command:

python3 shuffle_filenames.py -d "/path/to/dataset/dir" -e 2 -r "train-[0-9]+.jpg"

• -d: path to the directory where the dataset files are stored
• -e: number of training epochs
• -r: regex that corresponds to the filename of the dataset files. If this option is not used, all
• files in /path/to/dataset/dir will be considered.

After running this script, the file /home_dir/prisma/filenames_list will be created.

The shuffled_filenames module also provides a shuffle_filenames (files_list, n_epochs) function that can be called from inside another Python script. This function creates the filenames_list with the filenames present in files_list, considering that n_epochs training epochs will be performed, and returns a list with the content of filenames_list.

Read data

After including libprisma.so in your C++ project and creating the filenames_list, you simply need to instantiate Prisma and evoke its read method to read the dataset files. Here is a simple example:

#include <iostream>
#include <prisma/prisma.h>

int main() {
  // Create Prisma instance
  Prisma prisma = Prisma();

  std::string filename = "/path/to/file.txt";
  int n = 256;
  off_t offset = 0;
  char res[n];

  // Read 256 bytes from file.txt, starting at offset 0
  // Store result in res
  ssize_t bytes_read = prisma.read (filename, res, n, offset);

  // Print result
  std::cout << res << std::endl;

  return 0;
}

For this example to run as expected, the file /home_dir/prisma/filenames_list previously created, must contain a single line with /path/to/file.txt.

Integrating Prisma with Deep Learning frameworks

TensorFlow. Please refer to prisma/tensorflow_integration for details on how to integrate Prisma with TensorFlow.

PyTorch. Please refer to prisma/pytorch_integration for details on how to integrate Prisma with PyTorch.

Configurations

The number of I/O threads and buffer size used by Prisma can be configured using the configs file. This file must be stored in /home_dir/prisma when Prisma is executed.

Acknowledgments

We thank the National Institute of Advanced Industrial Science and Technologies (AIST) for providing access to computational resources of AI Bridging Cloud Infrastructure (ABCI). This work was financed by National Funds through the Portuguese Foundation for Science and Technology (FCT) on the scope of the UT Austin Portugal Program within project PAStor (UTA-EXPL/CA/0075/2019) and PhD Fellowships SFRH/BD/146059/2019 and SFRH/BD/146528/2019.

Contact

Please contact us at claudia.s.mendonca@inesctec.pt, rgmacedo@inesctec.pt or jtpaulo@inesctec.pt with any questions.