What It Actually Takes to Run Code on a €200 Million Supercomputer

MareNostrum V at the Universitat Politècnica de Catalunya in Barcelona is a €200 million public research infrastructure and one of the world's top 15 supercomputers. Rather than a single computer, it is a distributed computing environment where thousands of independent nodes are connected via an InfiniBand NDR200 fat-tree topology, guaranteeing non-blocking bandwidth.

Main Compute Partitions

• General Purpose Partition (GPP): Designed for highly parallel CPU workloads.
• Accelerated Partition (ACC): Equipped with NVIDIA H100 SXM GPUs for AI training and molecular dynamics.
• Login Nodes: Lightweight, task-only nodes for SSH access, file transfer, code compilation, and job submission.

Operations and Workload Management

• Airgap Environment: External internet access is blocked for security. This means pip install and similar commands are not possible, so all necessary libraries and datasets must be prepared in advance. Actual computation speed is so fast that extracting result data can itself become a bottleneck.
• SLURM Workload Manager: Users submit jobs to a queue via batch scripts, specifying the required number of nodes, number of tasks, execution time limits, and more. All jobs are subject to strict wall-time limits, and the process is terminated immediately upon exceeding them.
• Quantum Infrastructure: Digital-gate and superconducting-qubit-based quantum systems are integrated to implement hybrid classical-quantum computing, solving optimization problems that are difficult for classical supercomputers to handle.