Benchmarking Deep learning platforms by means of Tensorflow and Docker

From Lsdf

Description

The goal of the DEEP Hybrid DataCloud project [1] is to prepare a new generation of e-Infrastructures that harness latest generation technologies, supporting e.g. deep learning. In order to assess our solutions for new e-Infrastructures, a set of benchmark tools is needed.

Aim of this project is to prepare a set of benchmarks in form of Docker image(s) based on Tensorflow [2] and a number of pre-existing Tensorflow scripts (e.g. [3]), run them by means of various container tools and in different environments, compare performance.

Tasks

  • get acquanted with Docker containers and best practices [4, 5].
  • adapt existing Convolutional Neural Network (CNN) benchmarks [3] and pack them in Docker image(s) such that they can be run on GPUs in the DEEP testbeds.
  • run these benchmarks in HPC-like environment by means of udocker and singularity and on DEEP testbeds, use syntetic and real data, compare performance with bare-metal case.
  • (optional) extend the benchmarks for another neural network type (neural network code will be provided). Run same experiments as for CNN scripts.
  • at the end of Praktikum you write a short report with results of the performance tests (similar to [3]).

Requirements

  • programming experience in python
  • knowledge on how to install a program from source in Linux
  • ideally you have knowledge of containers and neural networks

Contact

valentin.kozlov@kit.edu

References

[1] https://deep-hybrid-datacloud.eu

[2] https://www.tensorflow.org

[3] https://www.tensorflow.org/performance/benchmarks

[4] https://www.docker.com

[5] https://docs.docker.com/develop/develop-images/dockerfile_best-practices/