Deployment of the computational environment for ML-INFN on INFN Cloud using GPUs (possibly partitioned)

INFN DataCloud Dashboard is a centralized portal that allows system administrators to instantiate services on the INFN Cloud federated infrastructure. The user has a portfolio of services available including the possibility of instantiating a virtual machine already configured for code development using Machine Learning algorithms; this service is called "Computational environment for Machine Learning INFN (ML_INFN)". Once the service has been selected, the user can enter information in the dedicated fields to customize its configuration, such as integration with IAM, the Jupyter image to use, whether to enable the resource monitoring system, mounting a CVMFS repository, deployment and partitioning of a GPU. This last feature, among others, is present in the new version of the service, and adds the possibility of partitioning specific models of Nvidia graphics cards, thus creating a greater number of virtual and less performing GPUs, usable by users of the configured instance of JupyterHub. This allows a user to overcome a previous limitation: multiple users could use JupyterLab instances at the same time only for applications requiring CPU usage or limited to the number of physical GPUs present on the machine. With partitioning, the maximum number of users who can simultaneously use a GPU, to be associated with their JupyterLab instance, has increased, at most, by a factor of 7, when considering the physical Nvidia A100 graphics card. Once the request for the service has been submitted, the portal will provide all the information necessary to access the virtual machine and the configured JupyterHub instance. This presentation was given in the ICT Infrastructures and Distributed Computing track of INFN CCR Workshop 2023