Sidebar
This is an old revision of the document!
Table of Contents
The Department of Computer Science hosts a computing cluster for scientific workloads. Please also consider the HPC Services offered by our ZID as they may suit your requirements better.
Hardware
The System runs in 3W06 in the ICT Building. It consists of 17 Nodes with the following specification
Motherboard: ASRock Fatal1ty X399 Professional Gaming CPU: AMD Ryzen Threadripper 2950X WOF Memory: G.Skill D4128GB 2400-15 Flare X K8 SSD: WD Black SN750 500 GB, NVMe 1.3 , Read: 3470 MB/s, Write: 2600 MB/s
Nodes gc1-gc8 offer
GPU: 4x ZOTAC GeForce RTX 2070 Blower, 8 GB (GDDR6, 256 Bit)
Nodes 9-16 offer
GPU: 4x ASUS GeForce RTX 2070 SUPER TURBO EVO, 8 GB (GDDR6, 256 Bit)
Nodes 17 offer
GPU: 1x NVIDIA TITAN RTX, 24 GB (GDDR6, 384 Bit)
The headnode (1x ASUS GeForce RTX 2070 SUPER TURBO EVO, 8 GB ) has 2TB additional storage capacity that is reachable from all nodes.
Network
The nodes and storage are interconnected through a 10 Gigabit Ethernet Switch. The upstream connection of the headnode is Gigabit Ethernet.
System Configuration
- All nodes run the latest Version of Ubuntu Server LTS which is 18.04.2 (Bionic Beaver).
- All nodes have the same home directories mounted.
- SLURM is used as job scheduler.
- ifi-auth is used as authentication backed.
Storage
Your home directory /home/name.surname should not exceed 100GB.Permission should be 700 in case you do not share it.
$chmod 700 /home/name.surname
A 18TB HDD space is available to the GPU Cluster as a scratch space. It uses the 1G network.
Mount point is : /scratch on each node
for better usage please:
create a directory for yourself using the same username you have for the cluster and set accessibility only to you.
(in case you need to share a directory give permissions for group/others 750 or 755)
$cd /scratch
$mkdir name.surname
$chmod -R 700 /scratch/name.surname
The IFI-NAS storage is mounted at /ifi-NAS/[your_group]. Information about the IFI_NAS is available here
Mounting an external resource:
first create your own directory
mkdir /mnt/users-mount/your_dir
them mount the external resource
sudo mount server:/resource /mnt/users-mount/your_dir
when finished, please do not forget to unmount it and delete the dirrectory:
sudo umount /mnt/users-mount/your_dir rm -rf /mnt/users-mount/your_dir
Installed Software Packages
Ubuntu Distribution Packages:
- Python 2.7
- Python 3.6
- gcc 7.4
- GNU Make 4.1
Loadable Modules:
- Open MPI 4.0.0 (/software-shared)
- mpich 3.3.1 (/software-shared)
Modules
Environment Modules can be used to modify a users environment. Use this to dynamically load modules in your sbatch script.
basic commands:
module avail -- list available modules module show [module name] -- show details about one module module help [module name] -- show help from one module
example usage in sbatch-script:
module load openmpi module unload openmpi
Authorization
To request access to ifi-cluster users have to subscribe to the mailing list.
Topology
To Be linked here soon.
Usage
After registration you can log into the system with your IFI credentials using ssh on ifi-cluster.uibk.ac.at.
Show running Jobs
squeue
Submit Job
Batch Job
write a configuration file similar to this basic example
- cluster-test.sh
#!/bin/bash -l #SBATCH --partition=IFIgpu #SBATCH --job-name=firstDemo #SBATCH --mail-type=BEGIN,END,FAIL ##optional #SBATCH --mail-user=max.mustermann@uibk.ac.at ##optional #SBATCH --account=your_group ##change to your group #SBATCH --uid=your_username ##change to your username #SBATCH --nodes=2 #SBATCH --ntasks-per-node=1 #SBATCH --mem=4G #SBATCH --time=0-00:30:00 #SBATCH --output slurm.%N.%j.out # STDOUT #SBATCH --error slurm.%N.%j.err # STDERR srun /bin/hostname | /usr/bin/sort
| Option | Description |
|---|---|
| --partition= | specifies on which partition you want to run your job. Available partitions are: IFIall-for CPU computation, IFIgpu -for GPU computation , IFItitan - for large GPU computation on a nVidia Titan GPU Card |
| --job-name= | you can define a name for your job (this field is optional) |
| --mail-type= | to receive an email if desired. Valid type values are NONE, BEGIN, END, FAIL, REQUEUE, ALL, +others see official documentation (this field is optional) |
| --mail-user= | specify the users email. it could be any UIBK email. if --mail-type is specified to other then NONE and --mail-user is missing emails will be send to local user on the IFI cluster (check it with “mail” command on headnode). For system wide mail forwarding please check here |
| --account= | specify the group you are belonging to |
| --uid= | specify your_username |
| --nodes= | on how many nodes should your job run. the scheduler will allocate the number of nodes to your job |
| --ntasks-per-node= | Request that ntasks be invoked on each node |
| --mem= | Specify the real memory required per node. Default units are megabytes. Different units can be specified using the suffix [K|M|G|T] |
| --time= | Set a limit on the total run time of the job allocation. If the requested time limit exceeds the partition's time limit, the job will be left in a PENDING state (possibly indefinitely) |
| --output=<filename pattern> | the batch script's standard output will be directly to the file name specified in the “filename pattern”. By default both standard output and standard error are directed to the same file |
| --error=<filename pattern> | the batch script's standard error will be directly to the file name specified in the “filename pattern”. |
srun /bin/hostname | /usr/bin/sort -- the command that you want to run
Run the file with sbatch --test-only <<filename>> to check the syntax and with sbatch «filename» to submit it.
Other valid and useful options:
| Option | Description |
|---|---|
| --exclusive[=user] | The job allocation can not share nodes with other running jobs (or just other users with the “=user” option). |
| --gres=<list> | Specifies a comma delimited list of generic consumable resources (GPUs). The format of each entry on the list is “name:type:count” (i.e - -gres=gpu:4 ) |
There is also an example script for C-Compilation.
Use the official Documentation website for details: https://slurm.schedmd.com/sbatch.html
Interactive Job
srun /bin/hostname
Interactive shell session
srun --nodes=1 --nodelist=gc3 --ntasks-per-node=1 --time=01:00:00 --pty bash -i
will get you an shell on gc3 node for one hour
Cancel Job
scancel <<jobid>>
The basic commands can be found in the SLURM User Guide
Live Monitoring Tool
The live monitoring tool for the cluster can be found atHERE
Contact
Contact ifi-sysadmin@informatik.uibk.ac.at if you have problem or further questions. For administration purposes there is also the internal documentation.
IFI Internal Chat
IFI Internal Chat Room for the GPU Cluster can be accessed at Matrix Chat
Last modified: 2021/03/25 10:04 by Mircea Munteanu