Grace Quick Start Guide
- 1 Grace Quick Start Guide
- 1.1 Deployment Status
- 1.2 Grace Usage Policies
- 1.3 Accessing Grace
- 1.4 Navigating Grace & Storage Quotas
- 1.5 Transferring Files
- 1.6 Managing Project Accounts
- 1.7 Finding Software
- 1.8 Running Your Program / Preparing a Job File
- 1.9 Submitting and Monitoring Jobs
- 1.10 tamubatch
- 1.11 Graphic User Interfaces (Visualization)
- 1.12 Deep Learning with TensorFlow and PyTorch
Cluster deployed, currently in testing and early user access mode.
Grace Usage Policies
Access to Grace is granted with the condition that you will understand and adhere to all TAMU HPRC and Grace-specific policies.
General policies can be found on the HPRC Policies page.
Most access to Grace is done via a secure shell session. In addition, two-factor authentication is required to login to any cluster.
Users on Windows computers use either PuTTY or MobaXterm. If MobaXterm works on your computer, it is usually easier to use. When starting an ssh session in PuTTY, choose the connection type 'SSH', select port 22, and then type the hostname 'grace.hprc.tamu.edu'. For MobaXterm, select 'Session', 'SSH', and then remote host 'grace.hprc.tamu.edu'. Check the box to specify username and type your NetID. After selecting 'Ok', you will be prompted for Duo Two Factor Authentication. For more detailed instructions, visit the Two Factor Authentication page.
Users on Mac and Linux/Unix should use whatever SSH-capable terminal is available on their system. The command to connect to Grace is as follows. Be sure to replace [NetID] with your TAMU NetID.
[user1@localhost ~]$ ssh [NetID]@grace.hprc.tamu.edu
Note: In this example [user1@localhost ~]$ represents the command prompt on your local machine.
Your login password is the same that used on Howdy. You will not see your password as your type it into the login prompt.
Off Campus Access
Please visit this page to find information on accessing Grace remotely.
For more detailed instructions on how to access our systems, please see the HPRC Access page.
When you first access Grace, you will be within your home directory. This directory has smaller storage quotas and should not be used for general purpose.
You can navigate to your home directory with the following command:
[NetID@grace1 ~]$ cd /home/NetID
Your scratch directory has more storage space than your home directory and is recommended for general purpose use. You can navigate to your scratch directory with the following command:
[NetID@grace1 ~]$ cd /scratch/user/NetID
You can navigate to scratch or home easily by using their respective environment variables.
Navigate to scratch with the following command:
[NetID@grace1 ~]$ cd $SCRATCH
Navigate to home with the following command:
[NetID@grace1 ~]$ cd $HOME
Your scratch directory is restricted to 1TB/250,000 files of storage. This storage quota is expandable upon request. A user's scratch directory is NOT backed up.
Your home directory is restricted to 10GB/10,000 files of storage. This storage quota is not expandable. A user's home directory is backed up on a nightly basis.
You can see the current status of your storage quotas with:
[NetID@grace1 ~]$ showquota
If you need a storage quota increase, please contact us with justification and the expected length of time that you will need the quota increase.
Files can be transferred to Grace using the scp command or a file transfer program.
Our users most commonly utilize:
- WinSCP - Straightforward, legacy
- FileZilla Client - Easy to use, additional features, available on most platforms
- MobaXterm Graphical SFTP - Included with MobaXterm
Advice: while GUIs are acceptable for file transfers, the cp and scp commands are much quicker and may significantly benefit your workflow.
Reliably Transferring Large Files
For files larger than several GB, you will want to consider the use of a more fault-tolerant utility such as rsync.
[NetID@grace1 ~]$ rsync -av [-z] localdir/ userid@remotesystem:/path/to/remotedir/
Managing Project Accounts
The batch system will charge SUs from the either the account specified in the job parameters, or from your default account (if this parameter is omitted). To avoid errors in SU billing, you can view your active accounts, and set your default account using the myproject command.
Software on Grace is loaded using hierarchical modules.
A list of the most popular software on our systems is available on the HPRC Available Software page.
To list all software installed as a module on Grace, use the mla utility:
[NetID@grace1 ~]$ mla
To search for a specific piece of software installed as a module on Grace using the mla utility:
[NetID@grace1 ~]$ mla keyword
To search for particular software by keyword, use:
[NetID@grace1 ~]$ module spider keyword
To see how to load a module, use the full module name:
[NetID@grace1 ~]$ module spider Perl/5.32.0
You will see a message like the following
You will need to load all module(s) on any one of the lines below before the "Perl/5.32.0" module is available to load. GCCcore/10.2.0
Load the base dependency module(s) first then the full module name
[NetID@grace1 ~]$ module load GCCcore/10.2.0 Perl/5.32.0
To list all currently loaded modules, use:
[NetID@grace1 ~]$ module list
To see what other modules can be loaded with the base dependency module (for example when GCCcore/10.2.0 is loaded)
[NetID@grace1 ~]$ module avail
To remove all currently loaded modules, use:
[NetID@grace1 ~]$ module purge
If you need new software or an update, please contact us with your request.
There are restrictions on what software we can install. There is also regularly a queue of requested software installations.
Please account for delays in your installation request timeline.
Running Your Program / Preparing a Job File
In order to properly run a program on Grace, you will need to create a job file and submit a job to the batch system. The batch system is a load distribution implementation that ensures convenient and fair use of a shared resource. Submitting jobs to a batch system allows a user to reserve specific resources with minimal interference to other users. All users are required to submit resource-intensive processing to the compute nodes through the batch system - attempting to circumvent the batch system is not allowed.
On Grace, Slurm is the batch system that provides job management. More information on Slurm can be found in the Grace Batch page.
The simple example job file below requests 1 core on 1 node with 2.5GB of RAM for 1.5 hours. Note that typical nodes on Grace have 48 cores with 384 GB of usable memory and ensure that your job requirements will fit within these restrictions. Any modules that need to be loaded or executable commands will replace the "#First Executable Line" in this example.
#!/bin/bash ##ENVIRONMENT SETTINGS; CHANGE WITH CAUTION #SBATCH --export=NONE #Do not propagate environment #SBATCH --get-user-env=L #Replicate login environment ##NECESSARY JOB SPECIFICATIONS #SBATCH --job-name=JobExample1 #Set the job name to "JobExample1" #SBATCH --time=01:30:00 #Set the wall clock limit to 1hr and 30min #SBATCH --ntasks=1 #Request 1 task #SBATCH --ntasks-per-node=1 #Request 1 task/core per node #SBATCH --mem=2560M #Request 2560MB (2.5GB) per node #SBATCH --output=Example1Out.%j #Send stdout/err to "Example1Out.[jobID]" #First Executable Line
Note: If your job file has been written on an older Mac or DOS workstation, you will need to use "dos2unix" to remove certain characters that interfere with parsing the script.
[NetID@grace1 ~]$ dos2unix MyJob.slurm
Submitting and Monitoring Jobs
Once you have your job file ready, it is time to submit your job. You can submit your job to slurm with the following command:
[NetID@grace1 ~]$ sbatch MyJob.slurm Submitted batch job 3606
After the job has been submitted, you are able to monitor it with several methods. To see the status of all of your jobs, use the following command:
[NetID@grace1 ~]$ squeue -u NetID JOBID NAME USER PARTITION NODES CPUS STATE TIME TIME_LEFT START_TIME REASON NODELIST 3606 myjob2 NetID short 1 3 RUNNING 0:30 00:10:30 2016-11-27T23:44:12 None tnxt-
To see the status of one job, use the following command, where XXXX is the JobID:
[NetID@grace1 ~]$ squeue --job XXXX JOBID NAME USER PARTITION NODES CPUS STATE TIME TIME_LEFT START_TIME REASON NODELIST XXXX myjob2 NetID short 1 3 RUNNING 0:30 00:10:30 2016-11-27T23:44:12 None tnxt-
To cancel a job, use the following command, where XXXX is the JobID:
[NetID@grace1 ~]$ scancel XXXX
tamubatch is an automatic batch job script that submits jobs for the user without the need of writing a batch script on the clusters. The user just needs to provide the executable commands in a text file and tamubatch will automatically submit the job to the cluster. There are flags that the user may specify which allows control over the parameters for the job submitted.
tamubatch is still in beta and has not been fully developed. Although there are still bugs and testing issues that are currently being worked on, tamubatch can already submit jobs to both the clusters if given a file of executable commands.
For more information, visit this page.
Graphic User Interfaces (Visualization)
The use of GUIs on Grace is a more complicated process than running non-interactive jobs or doing resource-light interactive processing.
You have two options for using GUIs on Grace.
The first option is to use the Open On Demand Portal, which is a web interface to our clusters. Users must be connected to the campus network either directly or via VPN to access the portal. More information can be found here, or on our YouTube channel
The second option is to run on the login node. When doing this, you must observe the fair-use policy of login node usage. Users commonly violate these policies by accident, resulting in terminated processes, confusion, and warnings from our admins.
Deep Learning with TensorFlow and PyTorch
Installing Python venv
# load all the required modules ml purge # CUDA modules are needed for TensorFlow ml GCCcore/9.3.0 GCC/9.3.0 Python/3.8.2 CUDAcore/11.0.2 CUDA/11.0.2 cuDNN/126.96.36.199-CUDA-11.0.2 # As Pytorch comes with CUDA libraries, we don't need to load CUDA modules. # the following two modules are sufficient for PyTorch # ml GCCcore/10.2.0 Python/3.8.6 # you can save your module list with (dl is an arbitrary name) module save dl # next time when you login you can simply run module restore dl # create a virtual environment (the name dlvenv is arbitrary) cd $SCRATCH python -m venv dlvenv source dlvenv/bin/activate
Installing Python packages
# First upgrade pip to avoid warning messages pip install -U pip # You can watch GPU usage on another terminal with watch -n 0.5 nvidia-smi
# install TensorFlow and other packages as needed. pip install tensorflow # Try things out (note that the login nodes grace4 and grace5 don't have GPUs) python -c "import tensorflow as tf; print(tf.test.gpu_device_name())"
pip install torch torchvision # Try things out (note that the login nodes grace4 and grace5 don't have GPUs) python -c "import torch; print(torch.device('cuda:0' if torch.cuda.is_available() else 'cpu'))"
All other python packages could be installed with pip install accordingly.
Sample Job Script - dl.slurm
#!/bin/bash ##ENVIRONMENT SETTINGS; CHANGE WITH CAUTION #SBATCH --export=NONE #Do not propagate environment #SBATCH --get-user-env=L #Replicate login environment ##NECESSARY JOB SPECIFICATIONS #SBATCH --job-name=JobExample4 #Set the job name to "JobExample4" #SBATCH --time=00:30:00 #Set the wall clock limit to 1hr and 30min #SBATCH --ntasks=1 #Request 1 task #SBATCH --mem=2560M #Request 2560MB (2.5GB) per task #SBATCH --output=Example4Out.%j #Send stdout/err to "Example4Out.[jobID]" #SBATCH --gres=gpu:1 #Request 1 GPU per node can be 1 or 2 #SBATCH --partition=gpu #Request the GPU partition/queue # modules needed for running DL jobs. Module restore will also work #module restore dl ml GCCcore/9.3.0 GCC/9.3.0 Python/3.8.2 CUDAcore/11.0.2 CUDA/11.0.2 cuDNN/188.8.131.52-CUDA-11.0.2 # Python venv source $SCRATCH/dlvenv/bin/activate # scripts or executables cd $SCRACTH/mywonderfulproject
Submit your slurm job with sbatch