Page Comparison

The Teton HPC cluster is the successor to of Mount Moran. Teton contains several new compute nodes. All Mount Moran nodes have been reprovisioned within the Teton HPC Cluster. The system is available by SSH using hostname teton.arcc.uwyo.edu or teton.uwyo.edu. We ask that everybody who uses ARCC resources cite the resources accordingly. See Citing Teton. Newcomers to research computing should also consider reading the Research Computing Quick Reference.

...

Global Filesystems

...

The Teton global parallel filesystem configured with a 160 TB SSD tier for active data and 1.2 PB HDD capacity tier for less-used data. The system policy engine moves data automatically between pools (disks and tiers). The system will automatically migrate data to HDD when the SSD tier reaches 70% used capacity. Teton has several spaces that are available for users to access described in the table below.

home/home/username ($HOME)

• Space for configuration files and software installations. This file space is intended to be small and always resides on SSDs. The /home file space is snapshotted to recover from accidental deletions.

project/project/project_name/[username]

• Space to collaborate among project members. Data here is persistent and is exempt from purge policy. No snapshots.

gscratch - /gscratch/username ($SCRATCH)

• Space to perform computing for individual users. Data here is subject to a purge policy defined below. Warning emails will be sent when possible deletions may start to occur. No snapshots.

Global Filesystems

Purge Policy -

• File spaces within the Teton cluster filesystem may be subject to a purge policy. The policy has not yet been defined. However, ARCC reserves the right to purge data in this area after 30 to 90 days of no access or from creation time. Before performing an actual purge event, the owner of the file(s) will be notified by email several times for files that are subject to being purged.

Storage Increases on Teton

...

Certain filesystems exist on different nodes of the cluster where specialized requirements exist. The table below summarizes these specialized filesystems.*Make a table!

Specialty Filesystems

The node-local scratch or lscratch filesystem is purged at the end of each job.

...

To request access for instructional use, send an email to arcc-info@uwyo.edu with the course number, section, and student list. If the PI prefers generic accounts can be created instead of providing a student list. Instructional accounts are usually valid for a single semester and access to the project is terminated at the beginning of the next semester.

System Access

...

SSH Access

Teton has login nodes for users to access the cluster. Login nodes are available publicly using the hostname teton.arcc.uwyo.edu or teton.uwyo.edu. SSH can be done natively on MacOS or Linux based operating systems using the terminal and the ssh command. Although X11 forwarding is supported, and if you need graphical support, we recommend using FastX if at all possible. Additionally, you may want to configure your OpenSSH client to support connection multiplexing if you require multiple terminal sessions. For those instances where you have unreliable network connectivity, you may want to use either tmux or screen once you login to keep sessions alive during disconnects. This will allow you to later reconnect to these sessions.

ssh USERNAME@teton.arcc.uwyo.edu

ssh -l USERNAME teton.arcc.uwyo.edu

ssh -Y -l USERNAME teton.arcc.uwyo.edu                          # For secure forwarding of X11 displays

ssh -X -l USERNAME teton.arcc.uwyo.edu                          # For forwarding of X11 displays

OpenSSH Configuration File (BSD,Linux,MacOS)

By default, the OpenSSH user configuration file is $HOME/.ssh/config which can be edited to enhance workflow. Since Teton uses round-robin DNS to provide access to two login nodes and requires two-factor authentication, it can be advantageous to add SSH multiplexing to your local environment to make sure subsequent connections are made to the same login node. This also provides a way to shorten up the hostname and access methods for SCP/SFTP/Rsync capabilities. An example entry looks like where USERNAME would be replaced by your actual UWYO username:

Host teton
  Hostname teton.arcc.uwyo.edu
  User USERNAME
  controlmaster auto
  controlpath ~/.ss/ssh-%r@%h:%p

WARNING: While ARCC allows SSH multiplexing, other research computing sites may not. Do not assume this will always work on systems not administered by ARCC.

Access from Microsoft Windows

ARCC currently recommends that users install MobaXterm to access the Teton cluster. It provides appropriate access to the system with SSH and SFTP capability, allowing X11 if required. The home version of MobaXterm should be sufficient. There is also PuTTY if a more minimal application is desired.

Additional options include, a Cygwin installation with SSH installed or the Windows Subsystem for Linux with an OpenSSH client installed on very recent versions of windows, enabling the OpenSSH client. Finally, a great alternative is to use our FastX capability.

FastX Access

If you’re currently on the UW campus, you can also leverage FastX to provide you with a more robust remote graphics capability via an installable client for Windows, Mac, or Linux or through a web browser. Navigate to https://fastx.arcc.uwyo.edu and log in with your 2FA credentials. There are also native clients for FastX for Windows, macOS, and Linux which can be downloaded here. For more information, see the documentation on using FastX.

Available Shells

Teton has several shells available for use. The default is bash]. To change your default shell, please submit the request through standard ARCC request methods.hard

GPUs and Accelerators

...

The ARCC Teton cluster has a number of compute nodes that contain GPUs. This section describes the hardware, as well as access and usage of the GPU nodes.

Teton GPU Hardware

The following tables list each node that has GPUs and the type of GPU installed.

Table #1

The following two GPU nodes are reserved for AI use.

Table #2

For more information on GPU capabilities please visit Nvidia's web page: [1]

For additional information about CUDA programming visit Nivida's CUDA C Programming Guide: [2]

Access and running jobs

There are three different types of GPU nodes in the Teton cluster and they are requested in somewhat different ways.

• Public Nodes: Public nodes are available to the general user and are in the "teton" partition. These nodes are identified in table #1 last column with a "yes". Use the following partition request to access these nodes.

--partition=teton

• Reserved Nodes: Reserved nodes are available to the general user and must be specifically requested via a partition request, i.e. "teton-gpu". These nodes are identified in table #1 last column with a "no". Use the following partition request to access these nodes.

--partition=teton-gpu

• Speciality Nodes: These are speciality nodes that are available to special users and are requested via a partition request, i.e. "dgx", see table #2 above. Use the following partition request to access these nodes.

--partition=dgx

If one wants to access the GPU devices on a node one MUST explicitly specify the generic consumable resources flag ("gres" flag). The "gres" flag has the following syntax:

--gres=<resource_type>:<resource_name>:<resource_count>

where:

• resource_type is always equal to gpu string for the GPU devices.

• resource_name is a string that describes the type of the requested gpu(s) e.g. k80, titanx, k20m, ....

• resource_count is the number of gpu devices that are requested of the type resource_name. Its value is an integer in the closed interval: {1,max. number of devices on a node}

The "gres" flag attached to each type of node can be found in the second-to-last column of Table 1. For example, the flag 

• -gres=gpu:titanx:1

 must be used to request one (1) GTX Titan X device that can only be satisfied by the nodes with the GTX Titan X in them.

If you run a job that requires GPUs and you fail to specify the "gres" flag, your job will be assigned any node in the requested partition. This means your job will possibly not have access to GPUs as part of your job.

One way to verify that your job has access to GPUs within a node you can execute the following command:

echo $CUDA_VISIBLE_DEVICES

An empty output string implies NO access to the node's GPU devices.

Some programs are serial, or able to run only on a single GPU; other jobs perform better on a single or small number of GPUs and therefore cannot efficiently make use of all of the GPUs on a single node. In order to better utilize our GPU nodes, node sharing has been enabled.. This allows multiple jobs to run on the same node, each job being assigned specific resources (number of cores, amount of memory, number of accelerators). The node resources are managed by the SLURM scheduler up to the maximum available on each node. It should be noted that while efforts are made to isolate jobs running on the same node, there are still many shared components in the system. Therefore a job's performance can be affected by other job(s) running on the node at the same time.

Node sharing can be accessed by requesting less than the full number of GPUs, CPUs or memory. Note that node sharing can also be done on the basis of the number of CPU's and/or memory, or all three. By default, each job gets 3.5 GB of memory per core requested (the lowest common denominator among our cluster nodes), therefore to request a different amount than the default amount of memory, you must use the "

• -mem

" flag . To request exclusive use of the node, use "

• -mem=0

".

Example #1

An example script that would request two Teton nodes with 2xK20m GPU's, including all cores and all memory, running one GPU per MPI task, would look like this:

#SBATCH --nodes=2
#SBATCH --mem=0
#SBATCH --partition=teton
#SBATCH --account=<account>
#SBATCH --gres=gpu:k20m:2
#SBATCH --time=1:00:00
... Other job prep
srun myprogram.exe

Example #2

To request all 8 K80 GPUs on a Teton node, again using one GPU per MPI task, we would do:

#SBATCH --nodes=1
#SBATCH --mem=0
#SBATCH --partition=teton
#SBATCH --account=<account>
#SBATCH --gres=gpu:k80:8
#SBATCH --time=1:00:00
... Other job prep
srun myprogram.exe

Example #3

Another example, using the job script below will get four GPUs, four CPU cores, and 8GB of memory. The remaining GPUs, CPUs, and memory will then be accessible for other jobs.

#SBATCH --ntasks=4
#SBATCH --nodes=1    
#SBATCH --mem=8
#SBATCH --partition=teton
#SBATCH --account=<account>
#SBATCH --gres=gpu:k80:4
#SBATCH --time=00:30:00 
... Other job prep
srun myprogram.exe

Example #4

To run a parallel interactive job with MPI, do not use the usual "srun" command, as this does not work properly with the "gres" request. Instead, use the "salloc" command, e.g.

salloc -n 1 -N 1 -t 1:00:00 -A <account> -p teton-gpu  --gres=gpu:p100:1

This will allocate the resources to the job, but keeps the prompt on the login node. You can then use "srun" or "mpirun" commands to launch the calculation on the allocated compute node resources.

For serial jobs, utilizing one or more GPUs, "srun" works properly, e.g.

srun -n 1 -N 1 -t 1:00:00 -A <account> -p teton-gpu --gres=gpu:p100:1 --pty /bin/bash -l

GPU programming environment

On Teton Nvidia CUDA, PGI CUDA Fortran and the OpenACC compilers are installed. The default CUDA is 9.2.88, which at the time of writing is the most recent. You can access by simply loading the CUDA module, "module load cuda". PGI compilers come with their own CUDA which is quite recent, and can be set access by loading the PGI module, using "module load pgi".

Any login node should work to compile your CUDA code as the CUDA tools are avaiable from the login nodes. PGI compilers come with their own CUDA so compiling anywhere from where you can load the PGI module should work.

To compile CUDA code using the CUDA compiler "nvcc" so that it runs on all types of GPUs that ARCC has, use the following compiler flags:

 -gencode arch=compute_35,code=sm_35 -gencode arch=compute_37,code=sm_37 -gencode arch=compute_52,code=sm_52 -gencode arch=compute_60,code=sm_60 -gencode arch=compute_70,code=sm_70

For more info on the CUDA compilation and linking flags, please have a look at [3].

The PGI compilers specify the GPU architecture with the -tp=tesla flag. If no further option is specified, the flag will generate code for all available computing capabilities (at the time of writing cc35,cc37, cc50, cc60, and cc70). To be specific for each GPU:

To invoke OpenACC, use the "

• acc

" flag. More information on OpenACC can be obtained at http://www.openacc.org.

A good tutorial on GPU programming is available at the CUDA Education and Training site from Nvidia.

When running the GPU code, it is worth checking the resources that the program is using, to ensure that the GPU is well utilized. For that, one can run the nvidia-smi command, and watch for the memory and CPU utilization. nvidia-smi is also useful to query and set various features of the GPU, see "nvidia-smi --help" for all the options that the command accepts.

For example, "nvidia-smi -L" lists the GPU card properties. On Teton node m025 you should see:

userX@m025:~# nvidia-smi -L
GPU 0: Tesla K20m (UUID: GPU-2e23ddef-1d96-7894-102a-0458da3faaa4)
GPU 1: Tesla K20m (UUID: GPU-458a86ec-09cd-64d1-475a-d36dc0a73b4f)

Debugging

Nvidia's CUDA distribution includes a terminal debugger named cuda-gdb. Its operation is similar to the GNU gdbdebugger. For details, see the cuda-gdb documentation.

For out of bounds and misaligned memory access errors, there is the cuda-memcheck tool. For details, see the cuda-memcheck documentation.

The Allinea DDT debugger that we currently license also supports CUDA and OpenACC debugging. Due to its user-friendly graphical interface we recommend them for GPU debugging. For information on how to use DDT or TotalView, see our debugging page.

Profiling

Profiling can be very useful in finding GPU code performance problems, for example, inefficient GPU utilization, use of shared memory, etc. Nvidia CUDA provides both command line (nprof) and visual profiler (nvvp). More information is in the CUDA profilers' documentation.