Connecting to the Clusters
Why can't I connect to the clusters from home?
You can but to do so requires passing via the EPFL VPN service. See http://network.epfl.ch/vpn for how to use this service.
Users preferring a command line tool might also wish to consider the tremplin SSH proxy tunnel service:. You can find the Linux and Windows procedure here.
Why am I asked for a password while sshing from the frontend to a node?
Once logged in to a frontend of a cluster, you can ssh directly to the node(s) running your job(s). You can prevent to be asked for the Gaspar password again by creating a passwordless ssh key.
Run the following command only once in any of the clusters:
Batch System Questions
What's the maximum run time of a job?
If you have a free account it's 6 hours. For paid accounts it's 3 days but you can ask to run for longer by contacting us and explaining why you need to run for more than 3 days.
How do I submit a job that requires a run time of more than three days?
Labs with signed contracts that have been granted permission to do so can add the "--qos=week" flag to ask for up to 7 days.
If your group has paid access then please contact SCITAS if you need such access.
Can I submit array jobs and, if so, how?
Can I have a paid and free account at the same time?
The short answer is no. Once a user is part of a group that pays for access (premium or share) then they are removed from the "free" group.
It may be that the group has only paid for access on one cluster so on the other clusters they will be part of the "free" group.
Is it safe to share nodes with other users?
Yes! We use cgroups to limit the amount of CPU and memory assigned to users. There is no way for users to adversely affect each other.
Is there a debug queue?
Not as such. In SLURM the concept of queues doesn't exist so to have priority access for debugging there is a partition which gives access to dedicated nodes:
The limits on the debug partition vary by cluster but in general the maximum run time is one hour and users are only allowed one job at a time. Interactive jobs are allowed.
I have a premium account and I have run on the debug partition. Do I have to pay for debug time?
No. Debug time is free of charge.
What is a <job id>?
It's the unique numerical identifier of a job and is given when you submit the job:
It can also be seen using squeue:
How do I display the used CPU time for my account since a certain point in time?
You can use the sreport tool. Here is an example of query where the used time is reported in core hours. Just replace 2018-01-01T00:00:00 with the start time you wish and scitas-ge with your account name.
How do I specify that my multi-node MPI job should use nodes connected to the same switch?
You can specify the maximum number of switches to be used as follows (in this case one switch)
Please note that jobs with such requirements may take much longer to schedule than those than can be spread across the cluster. This option should only be used in vey specific cases!
File System Questions
Where is my scratch space?
/scratch/<user name> - e.g.
Can you recover an important file that was on my scratch area?
**NO**. /scratch is not backed up so the file is gone forever. Please note that we automatically delete files on scratch to prevent it from filling up!
I've deleted a file on /home or /work - How can I recover it?
If it was deleted in the last seven days then you can use the daily snapshots to get it back. These can be found at:
/work/.snapshots/<date>/<laboratory or group>/
The home filesystem is backed up onto tape. If the file was deleted more than a week before we may be able to help. The /work filesystem is not backed up by default.
How to display scratch quota and usage information?
A. There are no quotas on scratch, as files older than 2 weeks may be deleted without notice as the filesystems fills up. However, you can see scratch usage for using the fsu command:
How to display quota and usage information for the /home and /work file systems?
A. /home: to get user quota and file system usage for your group members, use the following command:
You can also see an overview of /home usage and quota here.
B. /work: to get group quota and file system usage for your group members, use the following command:
You can also see an overview of /work file system usage and quota here.
How can I edit a file in the clusters using an application on my computer?
If you wish to manipulate file on the remote filesystem by using a software that is installed on your workstation
you can mount the remote filesystem by using sshfs. After install it, you can type from a terminal:
where USER is your gaspar account and MACHINE the cluster fs you wish to mount.
Why do I get the error "module: command not found" or "slmodules: command not found"?
This is probably because you have tcsh as your login shell and the environment isn't propagated to the compute nodes.
In order to fix the issue please change the first line of your job script as follows:
The -l option tells bash/tcsh to launch an interactive shell which correctly sources the files in /etc/profile.d/
Which options should I use to link with the Intel MKL?
If you use the Intel compilers then you can pass the -mkl flag which will do the hard work for you.
What compilers/MPI combination do you support?
SCITAS supports Intel compilers and Intel MPI (full proprietary) or GCC compilers and MVAPICH2 (full free) or GCC compilers and OPENMPI. Other combinations are not supported.
Why do some system tools stop working after the
intel module is loaded?
While this was true for the Cornalin (July 2017) release, it no longer applies as of the Paien (July 2018) release.
The Intel compiler includes its own versions of many libraries (and those take precedence over the system ones). Sometimes these libraries do not include symbols which are needed by some system tools and these will no longer work. Examples of these could be:
If a module exists providing the same tool one can just load that module.
If no module exists you will have to
module unload intel before using the command and
module load intel afterwards (any modules which depend on intel will simply become inactive and will be restored automatically).
Where should I compile code?
The best place to compile code is on the dedicated build nodes - see the build nodes section of the compiling codes documentation for more details.
Cluster / Partition specific FAQ
Why do I need to ask for access to the GPU nodes and how?
In order to use the GPU nodes we request that you submit a description of the code you wish to use and the performance benefits expected. You will then be invited to meet our application and GPU experts to discuss your proposal. This is in order to ensure that your code will make the best possible use of the resources and to make sure that you understand the features and limitations of the nodes. Non paying access is limited to a maximum run time of 12 hours and one task at a time.
Send an email to email@example.com with the title "HPC: Access to GPU"
How do I submit jobs to the GPU nodes?
If you pay for access then you need to pass the options "
--partition=gpu --qos=gpu --gres=gpu:X" where X is the number of GPUs per node required to sbatch.
If you have been granted "free" access then you need to pass the options "
--partition=gpu --qos=gpu_free --gres=gpu:X" where X is the number of GPUs per node required to sbatch.
How do I use one of the nodes with more than 64GB of memory?
Specify the amount of memory required with "--mem <quantity in MB>" either on the command line or in your job script.
How do I ask to use a specific processor type (Ivy Bridge, 16 cores or Haswell, 24 cores)?
For Ivy Bridge please give the option "
--constraint=E5v2" and for Haswell "
--constraint=E5v3". If you do not specify a constraint it may run on either but a multi-node job will never span both architectures.
How do I access the Deneb serial partition?
In order to replace Castor there is now a specific partition on Deneb for serial tasks that allows one to request individual cores.
To use this please pass the "--partition=serial” directive to SLURM:
Why did my job get sent to the serial partition?
Jobs that do not require a whole node are automatically routed to the serial partition at submission time so as to prevent resources being wasted. If you really do want your one CPU job to run on the parallel partition then pass the "
--exclusive" option to sbatch.
How many nodes are there in Fidis and what are their characteristics ?
There are 336 nodes with 128 GB of memory and 72 nodes with 256 GB of memory.
Each node have two Intel Broadwell processors running at 2.6 GHz. Each processor has 14 cores which makes 28 cores per node. A 800 GB local SSD disk (with 200GB allocated to /tmp) makes local checkpoints very fast.
All 408 nodes are interconnected with a non-blocking FDR infiniband network.
The Gacrux extension consists of 216 nodes, each with 192 GB of memory.
Each node has two Intel Skylake processors (Xeon 6132) each with 14 cores @ 2.6 GHz. The Gacrux nodes are interconnected with EDR Infiniband interconnect.
The nodes are arranged in non-blocking groups of 24 and the Skylake architecture offers significantly increased memory bandwidth.
Please note that to make use of the new AVX-512 instructions your codes will need to be recompiled. The centrally provided codes and libraries available through modules have been optimised for the new architecture.
How can I specify a particular type of node (Fidis or Gacrux)?
If you wish to specifically ask for the Gacrux/Skylake nodes then please use the following SLURM directive:
If you wish to use only the Fidis/Broadwell nodes then please specify:
If you do not specify a constraint then jobs may run on either partition but they will never span different architectures.
Are there debug nodes?
Two of the Gacrux nodes are available through the debug partition along with four Fidis nodes.
Why can't I connect to <random port>@<external server> from Fidis?
Fidis' network topology does not allow for direct connections to servers outside the EPFL network. Connections must always be done through an proxy server. This is already configured on all nodes for the most common use cases (ssh, git, http, https, ftp). For some cases you need to use specific versions of some tools: for example, to download a file from an FTP server (and if you can't use the Data Transfer Nodes) you need to use a more recent version of
curl (by doing
module load curl) and use the
If you find an use case which you depend on and does not work out of the box please contact us.