Driving Computational Efficiency in Large-Scale Platforms using HPC Technologies

TL;DR

The paper addresses how to improve HPC efficiency for large-scale astroparticle simulations in the LAGO project by analyzing historical job accounting data from the EGI FedCloud and classifying workloads into MC simulations, data processing, and user analysis/testing. It adopts a framework that combines logs with expert workflow knowledge to compute per-task efficiency metrics, such as CPU efficiency $\mathrm{CPU\ Efficiency} = \frac{Actual\ CPU\ Time}{Allocated\ CPU\ Time}$ and walltime efficiency $\mathrm{Walltime\ Efficiency} = \frac{Actual\ Walltime}{Requested\ Walltime}$. Key findings show that MC tasks have high per-task CPU and walltime efficiency, while short test jobs and site heterogeneity distort global averages, with data processing and interactive analyses exhibiting lower efficiency and nontrivial failure rates due to walltime overruns and out-of-memory events. The work provides concrete recommendations to refine resource requests, optimize workflow management, and guide future efficiency improvements, thereby increasing computational throughput and sustainability of LAGO’s HPC investments. These insights enable more accurate planning and execution of large-scale simulations and data analysis across heterogeneous sites, strengthening the scientific throughput of LAGO.

Abstract

The Latin American Giant Observatory (LAGO) project utilizes extensive High-Performance Computing (HPC) resources for complex astroparticle physics simulations, making resource efficiency critical for scientific productivity and sustainability. This article presents a detailed analysis focused on quantifying and improving HPC resource utilization efficiency specifically within the LAGO computational environment. The core objective is to understand how LAGO's distinct computational workloads-characterized by a prevalent coarse-grained, task-parallel execution model-consume resources in practice. To achieve this, we analyze historical job accounting data from the EGI FedCloud platform, identifying primary workload categories (Monte Carlo simulations, data processing, user analysis/testing) and evaluating their performance using key efficiency metrics (CPU utilization, walltime utilization, and I/O patterns). Our analysis reveals significant patterns, including high CPU efficiency within individual simulation tasks contrasted with the distorting impact of short test jobs on aggregate metrics. This work pinpoints specific inefficiencies and provides data-driven insights into LAGO's HPC usage. The findings directly inform recommendations for optimizing resource requests, refining workflow management strategies, and guiding future efforts to enhance computational throughput, ultimately maximizing the scientific return from LAGO's HPC investments.

Figures (5)

• Figure 1: Map of LAGO Detector Sites across Latin America, illustrating the geographical distribution and collaborative network contributing to astroparticle physics research. Source: Rubio_Montero_2021
• Figure 2: Overview of the LAGO Simulation Architecture. Source: Rubio_Montero_2021
• Figure 3: Conceptual Workflow for Analyzing HPC Resource Efficiency in the LAGO Framework, highlighting the steps from objective definition and data collection to workload characterization, efficiency analysis, and identification of optimization strategies.
• Figure 4: Examples illustrating the variability in execution times per task or particle within a single simulation job. These differences reflect the complexity of individual particle simulations, which are independent and submitted as separate tasks.
• Figure 5: Examples showing the distribution of simulation jobs in time. These graphs demonstrate the task-parallel execution strategy in which only the resources required by each task are requested, without idle reservations or inter-task dependencies.