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

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

• ZeroMQ with C++ binding
• Boost 1.66
• protobuf 3.4.1
• gperftools 2.7 (if build with TCMalloc)
• nlohmann-json
• concurrentqueue
• docopt.cpp
• easyloggingpp

See toplevel CMakeLists.txt for details.