Informing users about the new GPU resource management

Dear User!

We would like to draw your attention to some important changes regarding the management of the GPU resources in HUN-REN Cloud.

The popularity of HUN-REN Cloud has grown steadily in recent years, and GPU resources have been a major focus of interest. However, as a side effect of this welcome development, the available GPU capacity was essentially exhausted by this autumn, necessitating restructuring our current, rather flexible resource management system. More so, as our measurements show in many research projects the actual user utilization of the GPU resources they have reserved is uneven or relatively low, while the GPU demands of new users are not met at all or only after a long wait.

For these reasons, we had to change the GPU request and allocation rules. The goal is to implement a fairer GPU capacity allocation scheme that will make cloud services available to more users in a year, and also result in more efficient GPU resource utilization through the SLURM Job Manager.

From 1 January 2025, the new GPU Resource Management Policy for the HUN-REN Cloud will be applicable. We have developed a new GPU resource allocation scheme that will work exactly as the old scheme if sufficient resources are available. However, if there are insufficient resources, it will automatically switch to a mechanism that handles the competitive situation fairly.

Key changes (for IaaS projects that also use GPU resources), without being exhaustive:

  • Introduction of project phases
    • Preparatory project
      • Classification of new projects by default
      • Smallest flavor will be allocated
      • A learning phase with the possibility of consultation
    • Advanced project
      • Next project status after a successful preparatory phase
      • Larger flavor can be distributed
      • The preparatory phase can only be skipped if there is credibly proven GPU experience
  • Predefined run time
    • GPU resources can be requested for up to 3 months in the future, in special cases (service project or umbrella project) up to 1 year
  • For GPU projects already running
    • Projects expiring before 31.03.2025 will be closed on the original date
    • For projects expiring after 31.03.2025, the end date will change to 31 March 2025, but an extension request can be submitted until 1 March as follows:
      • You can apply for an extension of up to 3 months
      • Due to competing demands, a project may temporarily lack access to GPU resources
      • In case of an approved extension, the project can continue unchanged
      • No limit on the number of extensions a project can request
  • Evaluation and project run periods
    • Four-monthly system
    • Evaluation periods
      • Start: 1 March, 1 June, 1 September, 20 November.
      • End: on the first working day after the 14th day and at which time the applicants are notified
    • Project run periods
      • 1 Jan. - 31 Mar.
      • 1 Apr. - 30 Jun.
      • 1 Jul. - 30 Sep.
      • 1 Oct. - 31 Dec.
  • Requests for new and resource expansion
    • Submission as needed, continuously
    • The deadline for submission is before the start of the evaluation periods
    • They are evaluated immediately if resources are available, but typically during the next evaluation period
    • Requests received during the current evaluation period will only be considered in the following evaluation period if there are no resources available.
  • Evaluation criteria
    • Scientific value
    • Available resources
    • In case of an extension request, the results achieved with the project so far
    • New and extension requests are equally weighted
  • Start-up needs for new projects between 1 Jan. and 31 Mar. 2025
    • The project launch request must be accompanied by a completed GPU application form
    • A decision on the launch of projects will be made during the evaluation period starting on 1 March
    • Approved projects can start on 1 April
  • Extention requests for ongoing projects between 1 Jan. and 31 Mar. 2025
    • All projects wishing to continue their activities must register an extension request on the HUN-REN Cloud website by 28 Feb. 2025.
    • The request is submitted in the usual way but must be submitted together with a completed GPU application form
    • A decision on the continuation of projects will be made during the evaluation period starting on 1 March
    • Unauthorised projects must move off the HUN-REN Cloud by 31 March
    • Approved projects can continue to operate without interruption for another 3 months after 31 March
  • GPU request form structure
    • Which of the four types of IaaS projects you are requesting:
      • 3-month preparatory project
      • 3-month advanced project
      • 1-year service project
      • 1-year umbrella project
    • A brief description of the aim of the research and its expected scientific impact (max 1500 characters).
    • A list of the project members involved in the project who will be actually using the cloud, with their academic classification (PhD, habilitated associate professor, professor, doctorate, academician). Within this, the project leader (who is also the applicant) should be highlighted.
    • The expected percentage utilisation of the required GPU resources during the project.
    • The expected timing pattern of GPU usage (e.g. continuous, intermittent, how many usage periods and how many breaks between them, etc.) (max 1000 characters).
    • The name of international projects or organisations closely related to the project, a description of how they use the results of the project to be launched on the cloud or how they are otherwise related to the project (max 1500 characters).
    • List of publications undertaken (D1, Q1, Q2, presentations, etc.)
    • If the Cloud users involved in the project have other running or completed projects using GPUs on the HUN-REN Cloud or other Clouds, an itemized list of the results (publications, presentations, etc.) achieved in the project(s) (in such a case, the preparatory project phase can be skipped and an advanced project can be requested immediately).
    • In case of an extension request:
      • Is there a change in resource requirements (reduction or increase)?
      • Results achieved in the current 3-month project period (max 1000 characters).
    • Declaration:
      I declare that I will use the project running on the HUN-REN Cloud for the research described in the requested project within the institution specified.

New: SLURM Job manager

  • PaaS (Platform as a Service) to serve GPU needs
  • Ideal for running shorter, compute-intensive tasks with higher GPU capacity
  • FIFO (First In First Out) service scheduling and max 1 week run time
  • Launch expected from the second quarter of the year

Minimal changes will be made for projects that do not use GPUs at all: for these, new requests will continue to be accepted as usual and there will be no restrictions on extensions, but the maximum project duration that can be requested will be 1 year.

You can read more about this in the HUN-REN Cloud GPU Resource Management Policy to be published in January. Please read it carefully and feel free to contact us at info@science-cloud.hu with any questions you may have, and we will be happy to discuss further.

In January and February we will organise an online information event to explain the details to users.