Skip to content

cailun01/Salus

 
 

Repository files navigation

Salus: Fine-Grained GPU Sharing Primitives for Deep Learning Applications.

pipeline status

Implementation of Salus (arXiv paper).

Must be used with a customized tensorflow at SymbioticLab/tensorflow-salus. Note that these two projects are tightly coupled together. Make sure to use latest commit from both projects.

Try it out

Using docker. You will need the nvidia-docker extension.

docker run --rm -it registry.gitlab.com/salus/salus

This will start a salus server listen at port 5501.

Then when creating tensorflow session, use zrpc://tcp://localhost:5501 as the session target.

Compile yourself

Requires CMake 3.10 and modern compiler with c++14 support, e.g. GCC 5.4 is minimum.

Dependencies

See toplevel CMakeLists.txt for details.

About

Fine-grained GPU sharing primitives

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Jupyter Notebook 67.2%
  • Python 16.2%
  • C++ 15.3%
  • Other 1.3%