juliagpuplayground

For those who want to use Julia with GPU

Setup

• If you've installed nvidia-docker2, you can skip this chapter
• We will assume we are using a machine Ubuntu 20.04 with GPU e.g. GTX1080 or GTX1080-ti

```console $ lsb_release -a No LSB modules are available. Distributor ID: Ubuntu Description: Ubuntu 20.04.3 LTS Release: 20.04 Codename: focal $ nvidia-smi Sun Sep 12 09:32:22 2021
+-----------------------------------------------------------------------------+ | NVIDIA-SMI 470.63.01 Driver Version: 470.63.01 CUDA Version: 11.4 | |-------------------------------+----------------------+----------------------+ | GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |===============================+======================+======================| | 0 NVIDIA GeForce ... Off | 00000000:03:00.0 Off | N/A | | 0% 43C P8 11W / 280W | 8MiB / 11178MiB | 0% Default | | | | N/A | +-------------------------------+----------------------+----------------------+ | 1 NVIDIA GeForce ... Off | 00000000:04:00.0 Off | N/A | | 0% 45C P8 12W / 280W | 17MiB / 11177MiB | 0% Default | | | | N/A | +-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+ | Processes: | | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=============================================================================| +-----------------------------------------------------------------------------+ ```

TL;DR

• For those who hesitate to read my English, here are what we should do
  • Install NVIDIA driver
  • Install NVIDIA Container Toolkit
  • Install Docker/Docker Compose
  • Install GNU Make
• That's all. You can skip this chapter.

Install drivers

console $ sudo ubuntu-drivers autoinstall

Install Docker

• Please follow this instruction
• After reading this instruction, you'll find the following command works:

$ sudo docker run --rm --gpus all nvidia/cuda:11.0-base nvidia-smi

Having trouble ?

Well, the following commands to reset environments might be helpful for you:

console $ sudo apt remove --purge "nvidia-*" -y && sudo apt autoremove $ sudo apt remove --purge "cuda-*" -y && sudo apt autoremove $ sudo apt remove --purge "libcudnn*" -y && sudo apt autoremove $ sudo apt remove --purge "libnvidia-*" -y && sudo apt autoremove

console $ sudo apt-get remove docker docker-engine docker.io containerd runc

set "default-runtime": "nvidia"

• Edit /etc/docker/daemon.json to set "default-runtime": "nvidia"
  • See https://stackoverflow.com/questions/59691207/docker-build-with-nvidia-runtime

How to use this repository

Building Docker image

console $ make

Initialize Docker containerd

``console $ docker-compose run --rm julia Creating juliagpu_julia_run ... done _ _ _ _(_)_ | Documentation: https://docs.julialang.org (_) | (_) (_) | _ _ _| |_ __ _ | Type "?" for help, "]?" for Pkg help. | | | | | | |/ _ | | | | || | | | (| | | Version 1.6.2 (2021-07-14) / |_'|||_'| | Official https://julialang.org/ release |_/ |

julia> using Flux

julia> x = rand(10) |> gpu 10-element CUDA.CuArray{Float32, 1, CUDA.Mem.DeviceBuffer}: 0.481513 0.8270318 0.27988738 0.6559783 0.6754821 0.084349565 0.005481618 0.3933596 0.4915605 0.112065755 ```

Initialize JupyterLab

```console $ docker-compose up lab Starting gpujl-lab ... done Attaching to gpujl-lab

some stuff happen

gpujl-lab | [I 2021-09-12 09:50:20.717 ServerApp] jupyternbextensionsconfigurator | extension was found and enabled by nbclassic. Consider moving the extension gpujl-lab | [I 2021-09-12 09:50:21.040 ServerApp] Jupyter Server 1.11.0 is running at: gpujl-lab | [I 2021-09-12 09:50:21.040 ServerApp] http://blahblah:8888/lab?token=xxxxxxxxxxxxxxxxx gpujl-lab | [I 2021-09-12 09:50:21.040 ServerApp] or http://127.0.0.1:8888/lab?token=xxxxxxxxxxxxxxxxx gpujl-lab | [I 2021-09-12 09:50:21.040 ServerApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation). gpujl-lab | [W 2021-09-12 09:50:21.043 ServerApp] No web browser found: could not locate runnable browser. gpujl-lab | [C 2021-09-12 09:50:21.044 ServerApp] gpujl-lab | gpujl-lab | To access the server, open this file in a browser: gpujl-lab | file:///home/jovyan/.local/share/jupyter/runtime/jpserver-1-open.html gpujl-lab | Or copy and paste one of these URLs: gpujl-lab | http://blahblah:8888/lab?token=xxxxxxxxxxxxxxxxx gpujl-lab | or http://127.0.0.1:8888/lab?token=xxxxxxxxxxxxxxxxx ```

Then open your browser and go to http://127.0.0.1:8888/lab?token=xxxxxxxxxxxxxxxxx