Curnagl

Kesako?

Curnagl (Romanche), or Chocard à bec jaune in French, is a sociable bird known for its acrobatic exploits and is found throughout the alpine region. More information is available here

It's also the name of the HPC cluster managed by the DCSR for the UNIL research community.

A concise description if you need to describe the cluster is:

Curnagl is a 96 node HPC cluster based on AMD Zen2/3 CPUs providing a total of 4608 compute cores and 54TB of memory. 8 machines are equipped with 2 A100 GPUs and all nodes have 100Gb/s HDR Infiniband and 100Gb/s Ethernet network connections in a fat-tree topology. The principal storage is a 2PB disk backed filesystem and a 150TB SSD based scratch system. Additionally all nodes have 1.6 TB local NVMe drives.

If you experience unexpected behaviour or need assistance please contact us via helpdesk@unil.ch starting the mail subject with DCSR Curnagl.

An introductory tutorial video on using HPC clusters is available here.

How to connect

For full details on how to connect using SSH please read the documentation.

Please be aware that you must be connected to the VPN if you are not on the campus network. Then simply ssh username@curnagl.dcsr.unil.ch where username is your UNIL account.

You can also use the cluster through the OpenOnDemand interface.

Hardware

Compute

The cluster is composed of 96 compute nodes:

72 nodes with 2 AMD Epyc2 7402

24 nodes with 2 AMD Epyc3 7443

18 NVIDIA A100 (40 GB VRAM) distributed on 8 nodes

1 node with 2 AMD Epyc 9334 32-Core Processor and 8 NVIDIA L40S (48 GB VRAM)

1 NVIDIA GH200 (80 GB VRAM)

12 nodes with 1024 GB of memory, 512 GB otherwise.

Network

The nodes are connected with both HDR Infiniband and 100 Gb Ethernet. The Infiniband is the primary interconnect for storage and inter-node communication.

Cluster partitions

There are 3 main partitions on the cluster:

interactive

The interactive partition allows rapid access to resources but comes with a number of restrictions, the main ones being:

Only one job per user at a time

Maximum run time of 8 hours but this decreases if you ask for lots of resources.

For example:

CPU cores requested	Memory requested	GPUs requested	Run Time Allowed (h)
4	32	1	8
8	64	1	4
16	128	1	2
32	256	1	1

We recommend that users access this using the Sinteractive command. This partition should also be used for compiling codes.

This partition can also be accessed using the following sbatch directive:

#SBATCH -p interactive

There is one node with GPUs in the interactive partition and in order to allow multiple users to work at the same time these A100 cards have been partitioned into 2 instances each with 20GB of memory for a total of 4 GPUs. The maximum time limit for requesting a GPU is 8 hours with the CPU and memory limits applying. For longer jobs and to have whole A100 GPUs please submit batch jobs to the gpu partition.

Please do not block resources if you are not using them as this prevents other people from working.

If you request too many resources then you will see the following error:

salloc: error: QOSMaxCpuMinutesPerJobLimit`  
salloc: error: Job submit/allocate failed: Job violates accounting/QOS policy (job submit limit, user's size and/or time limits)`

Please reduce either the time or the cpu / memory / gpu requested.

cpu

This is the main partition and includes the majority of the compute nodes. Interactive jobs are not permitted. The partition is configured to prevent long running jobs from using all available resources and to allow multi-node jobs to start within a reasonable delay.

The limits are:

Normal jobs (3 days walltime):
- Maximun number of jobs submitted: 10000
- Maximum number of jobs running: 1152
- Maximum number of CPUs (used by all jobs same user): 1152
- Maximum number of memory (used by all jobs same user): 12T

Short jobs (12 hours):
- Maximun number of jobs submitted: 10000
- Maximum number of jobs running: 512
- Maximum number of CPUs (used by all jobs): 1536

Normal jobs are restricted to ~2/3 of the resources which prevents the cluster being blocked by long running jobs.

In exceptional cases wall time extensions may be granted but for this you need to contact us with a justification before submitting your jobs!

The cpu partition is the default partition so there is no need to specify it but if you wish to do so then use the following sbatch directive

#SBATCH -p cpu

GPU partitions

To request resources in a gpu partition please use the following sbatch directives:

#SBATCH --partition=<GPU_PARTITION>
#SBATCH --gres=gpu:<N>

Replace <N> with the number of needed GPUs (typically 1), and <GPU_PARTITION> with the name of one of the following partitions:

A100

Parition name: gpu

Number of available nodes in the partition: 7

Node configuration:
- 2x AMD EPYC 7402 24-Core Processor (x86_64)
- 2x NVIDIA A100 GPU 40GB
- Memory (RAM): 500GB

Recommended usage:
- General-purpose AI training for small to medium models
- Deep learning and advanced machine learning workloads
- HPC jobs requiring a balanced and versatile GPU

L40

Parition name: gpu-l40

Number of available nodes in the partition: 1

Node configuration:
- 2x AMD EPYC 9334 32-Core Processor (x86_64)
- 8x NVIDIA L40S GPU 46GB
- Memory (RAM): 750GB

Recommended usage:
- High-performance AI inference
- Suitable for medium-size training when needed

H100

Parition name: gpu-h100

Number of available nodes in the partition: 2

Node configuration:
- 2x AMD EPYC 9334 32-Core Processor (x86_64)
- 4x NVIDIA H100 GPU 94GB
- Memory (RAM): 750GB

Recommended usage:
- HPC workloads requiring extremely high memory bandwidth
- Training large AI models (LLMs, diffusion, multimodal)
- Optimized for the most demanding AI and transformer-based workloads

GH200

These nodes are specific because they use the Grace‑Hopper superchip (CPU + GPU) based on an ARM architecture. This means that there is a coherent and high-bandwidth access to memory from all computing units, providing high-performance computing, but it also means that all the software needs to be compiled and possibly optimized for this architecture.

Parition name: gpu-gh

Number of available nodes in the partition: 2

Node configuration:
- 1x Neoverse-V2 72-Core Processor (aarch64)
- 1x NVIDIA GH200 GPU 96GB HBM3
- Memory (RAM, LPDDR5X): 480GB
  - Coherent memory, also accessible from the CPU cores

Recommended usage:
- Advanced HPC tasks benefiting from large unified CPU-GPU memory
- Scientific applications with high memory bandwidth requirements
- Hybrid CPU+GPU AI workloads

Comparison of GPUs

GPU	FP64/TF64 (TFLOPS)	FP32/TF32 (TFLOPS)	TP16/BF16 Tensor (TFLOPS)	FP8 / INT8 (TFLOPS/TOPS)	Memory bandwidth	TDP
A100 40GB	9.7/19.5	19.5/156	312	INT8: 624	1.6 TB/s	250 W
L40s	-/-	91.6/183	362	FP8/INT8: 733	864 GB/s	300 W
H100 SXM5 94GB	34/67	67/494	1979	FP8/INT8: 3958	3.36 TB/s	700 W
GH200	34/67	67/494	989.5	FP8/INT8: 1979	4 TB/s	1000 W

Software

For information on the DCSR software stack see the following link:

https://wiki.unil.ch/ci/books/high-performance-computing-hpc/page/dcsr-software-stack

Storage

The recommended place to store all important data is on the DCSR NAS which fulfils the UNIL requirement to have multiple copies. For more information please see the user guide

This storage is accessible from within the UNIL network using the SMB/CIFS protocol. It is also accessible on the cluster login node at /nas (see this guide)

The UNIL HPC clusters also have dedicated storage that is shared amongst the compute nodes but this is not, in general, accessible outside of the clusters except via file transfer protocols (scp).

This space is intended for active use by projects and is not a long term store.

Cluster filesystems

The cluster storage is based on the IBM Spectrum Scale (GFPS) parallel filesystem. There are two disk based filesystems (users and work) and one SSD based one (scratch). Whilst there is no backup the storage is reliable and resilient to disk failure.

The role of each filesystem as well as details of the data retention policy is given below.

How much space am I using?

The quotacheck command allows you to see the used and allocated space:

$quotacheck 
------------------------------------------user quota in G-------------------------------------------
Path                     Quota   Used    Avail   Use% | Quota_files  No_files      Use%
/users/cruiz1            50.00   17.78   32.22    36% | 195852       202400         97%
------------------------------------------work quotas in T------------------------------------------
Project                              Quota   Used    Avail   Use% | Quota_files  No_files      Use%
pi_rfabbret_100222-pr-g              3.00    2.11    0.89     70% | 7098428      9990000        71%
cours_hpc_100238-pr-g                0.19    0.00    0.19      2% | 69713        990000          7%
spackbuild_101441-pr-g               1.00    0.00    1.00      0% | 1            9990000         0%

Users

/users/<username>

This is your home directory and can be used for storing small amounts of data. The per user quota is 50 GB and 100,000 files.

There are daily snapshots kept for seven days in case of accidental file deletion. See here for more details.

Work

/work/FAC/FACULTY/INSTITUTE/PI/PROJECT>

The work space is for storing data that is being actively worked on as part of a research project. This space can, and should, be used for the installation of any research group specific software tools including python virtual environments.

Projects have quotas assigned and while we will not delete data in this space there is no backup so all critical data must also be kept on the DCSR NAS. This space is allocated per project and the quota can be increased on request by the PI as long as free space remains.

Scratch

/scratch/<username>

The scratch space is for intermediate files and the results of computations. There is no quota and the space is not charged for. You should think of it as temporary storage for a few weeks while running calculations.

In case of limited space files will be automatically deleted to free up space. The current policy is that if the usage reaches 90% files, starting with the oldest first, will be removed until the occupancy is reduced to 70%. No files newer than two weeks old will be removed.

There is a quota of 50% of the total space per user to prevent runaway jobs wreaking havoc

$TMPDIR

For certain types of calculation it can be useful to use the NVMe drive on the compute node. This has a capacity of ~600 GB and can be accessed inside a batch job by using the $TMPDIR variable.

At the end of the job this space is automatically purged.