Architecture for the Lisa Global Fit Distributed Computing

Abstract

The Laser Interferometer Space Antenna (LISA) mission represents one of the most ambitiousgravitational wave detection projects, requiring unprecedented computational capabilities for processing heavy workloads through GlobalFit algorithms. This thesis investigates the deployment ofthe LISA GlobalFit distributed computing architecture on Portugal’s national High-PerformanceComputing (HPC) infrastructure, specifically the Deucalion supercomputer, to enable Portugal’sparticipation in the LISA Distributed Data Processing Centre (DDPC) network and create a Portuguese Distributed Computing Center (DCC).

The research addresses the critical challenge and comprehensive analysis of integrating cloudnative scientific workflow orchestration systems with traditional HPC environments and the integration challenges between Singularity-based HPC environments and Kubernetes-based cloudnative architectures.

The LISA GlobalFit framework employs sophisticated Markov Chain Monte Carlo (MCMC)algorithms that require distributed computing across multiple international HPC centers, presenting unique architectural and technological compatibility challenges. The methodology adopteda phased approach, beginning with successfully validating the isolated GlobalFit framework onDeucalion’s heterogeneous ARM-based architecture, followed by attempts to deploy the DCCarchitecture.

Key findings demonstrate that Portugal has the technical capabilities to contribute meaningfully to LISA data processing. The isolated GlobalFit framework was successfully deployed,achieving 100% CPU utilization and validating critical checkpoint and resume functionality essential for long-running HPC computations. The framework successfully scaled from simpletest cases involving single Gaussian fitting to complex scenarios with thousands of sources andsamples, confirming its suitability for LISA-scale data processing requirements. However, the research revealed fundamental architectural incompatibilities between traditional HPC containerization technologies (Singularity) and modern cloud-native orchestration requirements (Kubernetesbased systems). The absence of OCI-compatible container engines on Deucalion created insurmountable barriers for deploying the DCC architecture, highlighting broader tensions betweentraditional HPC infrastructure and evolving cloud-native scientific computing paradigms.

The main contributions are: (1) a replicable methodology for deploying the GlobalFit framework on heterogeneous HPC systems; (2) documentation and architectural analysis filling key gapsin existing resources; (3) a technical assessment of challenges and solutions in multi-architecturedeployment; and (4) strategic insights for the EuroHPC community on integrating traditional HPCwith cloud-native scientific workflows.

The research establishes Portugal’s technical readiness for LISA participation while identifying critical infrastructure requirements for full DDPC integration. Although DCC architecturedeployment was not achieved due to container engine incompatibilities, the work provides valuable insights that can inform future European HPC infrastructure decisions and contribute to thebroader understanding of HPC-cloud integration challenges in scientific computing.