Prerequisites

Please have these installed before proceeding

• Clang
• Protocol Buffers
• gRPC
• Ninja
• Meson >= 0.55
• TensorFlow < 2
• CUDA, NCCL

Build instructions

1. Configure the build:

   CXX=clang++ meson build
   

   This command configures the build in a directory named build.

   It is recommended that you use lld as the linker. To do so, use the following command instead:

   CXX=clang++ CXX_LD=lld meson build
   

2. Build eLF:

   ninja -C build
   

3. Do some tests

   meson test -C build -v
   

Usage

• Start elf.Controller
• Modify TensorFlow code with elf.Worker and elf.tensorflow.ElasticOptimizer
• Start worker code to scale out
• Call Controller.leave(worker_id) on the controller or Worker.leave() on the worker program to scale in