Usage policy

Additional information

General Terms of Use for CSC's Services for Research and Education

When you login to the CSC supercomputer, you end up on one of the login nodes of the cluster. These login nodes are shared by all users and they are not intended for heavy computing.

The login nodes should be used only for:

code editing and code compilation
managing batch jobs
moving data
light pre- and postprocessing

Here light means one-core jobs that finish in minutes and require less than 1 GiB of memory at maximum. All other tasks are to be done in compute nodes either as normal batch jobs or as interactive batch jobs. Programs not adhering to these rules will be terminated without warning.

Important

The login nodes are not meant for long or heavy processes.

Agentic AI tools on Roihu

Agentic AI tools, such as Claude Code, OpenAI Codex and OpenCode, are LLM-based assistants that can autonomously run commands, edit files and submit jobs on your behalf. They are typically used for tasks like coding assistance or managing Slurm jobs. AI agents should be used carefully, as they can introduce security risks and performance problems for all Roihu users, including yourself.

Responsibility for running AI agents

You are always responsible for the actions of your AI agents. Any command run by your agent is executed under your personal user account, and you are bound by the General Terms of Use regardless of whether you or your agent typed the command.

Key rules for using AI agent tools on Roihu:

Use up-to-date csc-skills. These skills provide guidance for the tools on how to use the system without causing disruptions, and help with the other points in this list.
Control the agent's access to files on the system. Do not give the agent access to files that contain your secrets, or any other information that you do not want to provide to it. Typical examples are your SSH keys in ~/.ssh and Allas credentials in rclone.conf. A good method is to run the agentic tool in an Apptainer container, only bind mounting what you need on the system (e.g. your scratch directory). CSC is preparing a supported container for this use case. One can also use the sandboxing features of the tools, but not all methods are supported on Roihu. See for example Claude Code sandbox environments.
Control the agent's access to tools on the system. Be mindful of what tools you allow the agent to execute without confirmation. The more isolated the sandbox, the more permissive you can afford to be with auto-approved tools. See the tool documentation on how to configure this, for example Claude Code permissions and Codex sandbox and approval policies.
Prefer running the agent on your own computer. If possible, run the tool on your own laptop or workstation and not on the supercomputer. The agent can then run commands and manage jobs through the FirecREST API. Keep the FirecREST token in an environment variable (e.g. FIRECREST_TOKEN), let the agent refer to it by name only, never paste it into a prompt and do not let the agent print it. CSC is preparing a FirecREST MCP server that keeps the token outside the agent's reach. Roihu can also be accessed over SSH.
Do not overload the login node! As an exception to the login node rules above, running the interactive session of an agent on a login node is allowed, provided that you supervise it and that the agent itself only does light work. All heavy computation must be submitted as normal batch jobs or interactive batch jobs.
Do not overload Lustre and Slurm! If you run the agent on Roihu, instruct it that home, projappl and scratch are on a Lustre filesystem and that it should use the local disk under $TMPDIR for temporary files. Avoid excessive amounts of Slurm jobs and heavy tool calling: no tight polling loops of squeue or sacct, no recursive find or grep over large directories on Lustre, and no bursts of test jobs.
Never give your CSC credentials to an agent running on a third-party service. Do not give your CSC password, SSH keys or FirecREST tokens to an agent hosted on a system that you do not control, such as a web-based chatbot with tool calling or a cloud-based AI-assisted IDE. Under the General Terms of Use you agree not to share your credentials or leave them for others to see, and this includes handing them to a service you do not control. It is permitted to run an agent on your own computer where your SSH keys are stored.
Do not open Roihu to an agent running on a third-party service. Do not start an MCP server, a reverse tunnel or a similar service on Roihu that lets an agent running on a system you do not control execute commands on Roihu. This gives that service the same access to your account as sharing your credentials would.
Do not expose other users' data to the agent. On a shared system you can see information about other users that is not yours to pass on: their user names, jobs and processes on login nodes. Everything the agent reads may be sent to an external LLM provider, so restrict the agent to your own data. For example, always list only your own jobs (squeue --me) and processes (ps -u $USER), and do not let the agent run commands like squeue, sacct -a, who, top or ps aux without restriction.

For a more extensive list of common problems with AI agents and how to avoid them, see the LUMI AI agent guide. The same principles apply on Roihu.

Disk cleaning

Each project has disk space in the directory /scratch/<project>. This fast parallel scratch space is intended for data that is in active use. To ensure that the parallel disk system does not run out of storage space and to keep performance acceptable, CSC automatically removes files in Roihu scratch that have not been accessed in a long time. The performance of a parallel file system starts to degrade when it fills up, and the more it fills up, the slower the performance will get.

This cleaning will happen regularly, and each time users are informed at least 1 month in advance. CSC also provides lists of files that are about to be removed and instructions for how one can transfer important files to more suitable disk systems.

The cleaning is stricter for projects with larger quotas:

For projects that have a scratch quota of 5 TiB or more, files that have not been accessed (opened, read, modified) in the last 90 days will be deleted.
For other projects with smaller scratch quotas, files that have not been accessed (opened, read, modified) in the last 180 days will be deleted.

You can use the csc-workspaces command to see which cleaning cycle your projects are subject to.

GPU nodes

Roihu GPUs should be used for workloads that benefit from GPU capacity compared to using CPUs or which can't be run on CPUs. A good rule of thumb is to compare the Billing Unit (BU) usage (e.g. with seff or the Billing Unit calculator) of the job on GPUs against CPUs and select the one using less. One CPU BU and one GPU BU are equal in terms of cost.

LUMI has a lot of GPU capacity which is also "cheaper" as measured in BUs, and on LUMI it's better to use GPUs if possible for your research. In any case, always make sure you use resources efficiently.

Conda installations

Due to performance issues of Conda-based environments on parallel file systems, CSC has deprecated the direct usage of Conda installations. This means that any Conda environments you intend to use must be installed within a container. See Conda best practices for more information.

Tykky

Please consider the Tykky container wrapper for easy containerization of Conda and pip environments.

Running out of Billing Units

When a project runs out of Billing Units, the ability to use the service will be limited in three phases. If you are still actively using the project you can lift the limitations by applying for more Billing Units.

In the first phase the ability to submit new jobs is limited:

If you run out of Storage BUs, no new jobs can be submitted to any partition
If you run out of CPU BUs, no new jobs can be submitted to CPU partitions
If you run out of GPU BUs, no new jobs can be submitted to GPU partitions

In other words, running out of CPU or GPU BUs only affects the corresponding partition type, while Storage BUs affect all. Jobs that are running are not interrupted and will run until completion/timeout.

In the second step data access is limited. When you run out of storage BUs a 30-day grace period starts, after which access to /projappl and /scratch folders is disabled. No data is deleted, it is only access that is disabled. Data will, however, still be removed from /scratch during the normal cleaning process. Note that having negative balance for CPU or GPU BUs does not trigger this step, only a negative Storage BU balance.

If you are not using a project actively we encourage you to migrate any data that you still need within the 30-day grace period and then close the project in MyCSC.

In the third phase the project is closed after a 60-day grace period if you have run out of BUs of any type. If the project still has a negative amount of Billing Units of any type after 60 days, it will be closed.

Slurm job management by CSC

CSC will not change job parameters like length or priority.
CSC can terminate jobs if they are misusing resources. E.g., if resources (CPU cores, GPUs, memory) are severely underutilized or IO is overloading the storage system.

Robot accounts on Roihu

The following policies should be followed when applying for machine-to-machine robot account on Roihu. Robot accounts are intended for automation (e.g. automatic data processing pipeline obeying rules below) and integrations to external services or infrastructures.

Any use where the Robot account is used as a shared account is considered account sharing, which is against the General Terms of Use.

Rules concerning acount creation, management and secrets:

Only CSC project managers and vice project managers can request a robot account.
Project manager and vice project managers are always responsible for the robot account and its use.
All usage and all uses of the solution using robot accounts must comply with the General Terms of Use, Usage policy (this page) and Export restrictions and sanctions.
Use of passwords and ssh-keys of personal user accounts in robot accounts is prohibited. This is considered sharing of credentials and is against Terms of Use.
Credentials associated with robot accounts, including but not limited to passwords, SSH private keys, API tokens, authentication tokens, cloud access keys, and any other authentication or authorization material, must be stored in a way that prevents access by unauthorized users. SSH keys should have a non-empty passphrase. We recommend using secrets management system, local credential managers or encrypted password managers for other credentials.
Credentials must never be transmitted over insecure channels (e.g. email, chat, unencrypted HTTP, ticketing systems), embedded in source code, committed to version control, written to logs, or included in container images, build artifacts, or error messages.

Rules concerning Roihu usage:

On Roihu, computing resources can be accessed with a robot account using the FirecREST HPC API. Direct SSH terminal login with a robot account is not available except in exceptional cases.
Inbound and outbound network traffic generated by robot accounts are allowed as long the the network traffic consists of data required by or generated by the project.
Users who are not members of the project associated with a robot account must not be able to execute any commands, code or API calls via the robot account.
Robot accounts may only execute applications, scripts, binaries and API calls explicitly defined by project members.
Robot accounts must not accept, interpret, or execute user-supplied code or commands that could alter execution flow. All inputs from external users must be treated as untrusted data, not instructions.
Services exposed via robot accounts must validate and sanitize all external inputs before processing.

CSC may audit and monitor the usage:

CSC reserves the right to audit and monitor usage of robot accounts. Should CSC find suspicious activity, CSC may close the robot account pending thorough review.
Should it be found that a robot account is being misused, CSC reserves the right to determine appropriate measures on case-by-case basis, including closure of project and user accounts involved in misuse.