Skip to content

Getting started on 8-way H100 nodes on Satori

This page provides information on how to run a first job on the IBM Watson AI Lab H100 GPU nodes on Satori. The page describes how to request an access to the Slurm partition associated with the H100 nodes and how to run a first example pytorch script on the systems.

A first set of H100 GPU systems has been added to Satori. These are for priority use by IBM Watson AI Lab research collaborators. They are also available for general opportunistic use when they are idle.

Currently ( 2023-06-19 ) there are 4 H100 systems installed. Each system has 8 H100 GPU cards, two Intel 8468 CPUs each with 48 physical cores and 1TiB of main memory.

Below are some instructions for getting started with these systems.

Access to the nodes

To access the nodes in the priority group you need your satori login id to be listed in the WebMoira group https://groups.mit.edu/webmoira/list/sched_oliva. Either Alex Andonian and Vincent Sitzmann are able to add accounts to the sched_oliva Moira list.

Interactive access through Slurm

To access an entire node through Slurm, the command below can be used from the satori login node

srun -p sched_oliva --gres=gpu:8 -N 1 --mem=0 -c 192 --time 1:00:00 --pty /bin/bash

this command will launch an interactive shell on one of the nodes (when a full node becomes available). From this shell the NVidia status command 

nvidia-smi
should list 8 H100 GPUs as available.

Single node, multi-gpu training examples (for example https://github.com/artidoro/qlora ) should run on all 8 GPUs.

To use a single GPU interactively the following command can be used 

srun -p sched_oliva --gres=gpu:1 --mem=128 -c 24 --time 1:00:00 --pty /bin/bash

this will request a single GPU. This request will allow other Slurm sessions to run on other GPUs simultaneously with this session.

Running a nightly build pytorch example with a fresh miniconda and pytorch

A miniconda environment can be used to run the latest nightly build pytorch code on these systems. To do this, first create a software install directory and install the needed pytorch software

mkdir -p /nobackup/users/${USER}/pytorch_h100_testing/conda_setup

and then switch your shell to that directory. 

cd /nobackup/users/${USER}/pytorch_h100_testing/conda_setup

now install miniconda and create an environment with the needed software 

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh 
chmod +x Miniconda3-latest-Linux-x86_64.sh
./Miniconda3-latest-Linux-x86_64.sh -b -p minic
. ./minic/bin/activate 
conda create -y -n pytorch_test python=3.10
conda activate pytorch_test                          
conda install -y -c conda-forge cupy
pip3 install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121

Once the software is installed, the following script can be used to test the installation.

test.py
import torch
device_id = torch.cuda.current_device()
gpu_properties = torch.cuda.get_device_properties(device_id)
print("Found %d GPUs available. Using GPU %d (%s) of compute capability %d.%d with "
          "%.1fGb total memory.\n" % 
          (torch.cuda.device_count(),
          device_id,
          gpu_properties.name,
          gpu_properties.major,
          gpu_properties.minor,
          gpu_properties.total_memory / 1e9))
Run this script with 
python test.py

To exit the Slurm srun session enter the command 

exit

Running a simple batch script using an installed miniconda environment

To run a batch script on one of the H100 nodes in partition sched_oliva first paste the content in the box below into a slurm script file called, for example, test_script.slurm ( change the RUNDIR setting to assign the path to a directory where you have already installed a conda environment in a sub-directory called minic ).

First create the mytest.py script with the contents above.

#!/bin/bash
#
#SBATCH --gres=gpu:8
#SBATCH --partition=sched_oliva
#SBATCH --time=1:00:00
#SBATCH --mem=0
#

nvidia-smi

RUNDIR=/nobackup/users/${USER}/h100-testing/minic
cd ${RUNDIR}

. ./minic/bin/activate

conda activate pytorch_test

python mytest.py

This script can then be submitted to Slurm to run in a background batch node using the command.

sbatch < test_script.slurm

Getting help

As always, please feel welcome to email orcd-help@mit.edu with questions, comments or suggestions. We would be happy to hear from you!