abcinje

Host

1. Install docker

$ sudo apt update
$ sudo apt install -y docker.io
$ sudo usermod -a -G docker $USER
$ sudo service docker restart
$ echo "Re-login required to start docker"
$ logout

abcinje

1. Generate a container

$ docker run -it --name bufferring --network=host nvidia/cuda:10.2-devel /bin/bash

2. Install packages

# apt update
# apt install -y build-essential git openssh-server python3 python3-pip vim wget

abcinje

Computation Thread

computation_thread()

for each epoch {
  for each step {
    for each parameter in propagation_process {
      grad = compute_gradient(parameter);
      handle = asynchronous_allreduce(grad);
      handles[grad] = handle;

abcinje

CkptFS

Running drainer

# 환경변수 CKPT, BB, PFS가 set되어 있는 경우
build/bin/drainer

# 그렇지 않은 경우
CKPT= BB= PFS= build/bin/drainer