Skip to content

Latest commit

 

History

History
81 lines (66 loc) · 3.36 KB

BIMK-Server-HandBook-en.md

File metadata and controls

81 lines (66 loc) · 3.36 KB
Raw

BIMK GPU Server Handbook V 1.0

ZH

Maintainer Last Update Time Status
@Anonymone 2021.04.17 Out-dated

This file contains all the basic notifications you need to know before you use the GPU server of BIMK.

Platform

BIMK has one GPU Server with 4 NVIDIA Tesla P100 GPU cards which run on the Ubuntu 18.04(bionic) Desktop release. However, the server does not support running any GPU-required programs directly on the low-level system layer. Instead, Singularity as a flexible virtual technique is used to eliminate the problems of software compatibility. Here is a description of why we use Singularity:

User A: need to run an old program that requires Pytorch 0.4.0 and python 3.5. 

User B: need to run a new program that requires Pytorch 1.4.0 and python 3.6.  
    ......

It is hard to manage all these requirements with effect due to the reason that some software can not exist together in the systems layer or it might be an unpractical approach. Fortunately, Singularity tackles this problem by making different software environments to be independent with each other.

Create Your Account

For security consideration, we do not provide the self-registered service. You need to contact the administrator.

After application and authorization, you can login the GPU server with ssh. The initial password is 123456 and you can change it with the following command:

passwd

CAUTION: We do not provide any root privilige for users. The key reason root in the fact that users with root privilage may tent to have an irreversible damage on GPU server.

Fast Book of Singularity

Many pre-build Sigularity images are provided at /Share/singularity/images and the names of them are organized as 'system-{software1-sofware2-...(option)}-build_date.sif'.

Use Pre-build Images and Run Your Programs

For instance, we have a program file hijack.py as follow:

#!/usr/bin/env python3
# encoding: utf-8
import torch 
print(torch.cuda.is_avvailable())

You can exec this script by using this command:

singularity exec ~/PATH/YOUR_IMAGE python3 hijack.py

Note that ~/PATH/YOUR_IMAGE is the path where the image you need and python3 hijack.py is the command you want to execute. If you encounter an error like no gpu device available but the server resources are sufficient, you can add the command --nv after exec to solve it.

Customize Your Image

In general cases, you may need to build your personal images. Singularity provides a simple approach to do it.

  1. Step1 Install the Singularity on your local device with root(Linux release) How to install the Singularity?

  2. Step2 Copy the Basic Image /Share/singularity/images/base-nvidia.sif to local.

scp /Share/singularity/images/base-nvidia.sif  YOUR_ACCOUNT@YOUR_LOCAL_IP:~/YOUR_LOCAL_PATH
  1. Step3 Create A Writable Image
sudo singularity build --sandbox ~/PATH_to_SAVE_IMAGE ~/YOUR_LOCAL_PATH/base-nvidia.sif
  1. Step4 Enter the Container of Image
sudo singularity shell --writable ~/PATH_to_SAVE_IMAGE
  1. Step5 Install All You Need as Usual Such as:
sudo apt update && \
sudo apt install python3-pip -y
  1. Step6 Exit the Container
exit
  1. Step7 Build Your Image
sudo singularity build ~/PATH_TO_SAVE/IMAGE_NAME.sif ~/PATH_to_SAVE_IMAGE