LRZ Demonstrates Scheduler-Managed QPU Integration on SuperMUC-NG Using MQSS and Slurm GRES

The arXiv preprint (v1, Aug 2026) details an open-source SLURM Plugin Suite of Prolog/Epilog scripts and SPANK modules that exposes QPUs as Generic RESources without altering the core scheduler. System-level allocation occurs via Slurm while the MQSS Quantum Resource Manager & Compiler Infrastructure handles just-in-time circuit compilation and dispatch. Experiments on the petascale system showed hybrid workloads executing alongside conventional MPI jobs with overhead below measurement noise. This architecture directly targets next-generation HPE Cray platforms such as LRZ's planned Blue Lion.

Prior hybrid efforts relied on ad-hoc API bridges or dedicated quantum partitions that fractured existing HPC scheduling policies. By contrast, the MQSS model reuses production resource managers, reducing the operational surface that must be validated for security and accounting. The 20-qubit scale remains modest, yet the integration pattern scales to larger QPUs once error rates permit useful circuit depths. Portability claims rest on the assumption that future Cray systems retain compatible Slurm GRES semantics.

Related work includes the 2024 IEEE paper on Qiskit Runtime integration with LSF at Oak Ridge and the 2025 arXiv preprint on XACC-Slurm co-scheduling at Sandia; both reported higher orchestration latency when bypassing native schedulers. The LRZ result therefore supplies the first quantified evidence that a standards-compliant plugin layer can eliminate that penalty at petascale. Remaining gaps include multi-tenant QPU sharing policies and fault-tolerant circuit recompilation under node failures.

Next steps involve porting the plugin suite to Blue Lion's Slurm configuration and benchmarking against 100-qubit targets expected in 2027. Success metrics should include job throughput, queue wait-time distributions, and energy-per-circuit rather than raw fidelity alone.