Article Preview
TopIntroduction
Separating High-Energy Physics (HEP) developments and experiments from computational approaches to data analysis is currently an infeasible task. For instance, the Large Hadron Collider (LHC) at CERN in Geneva produces several petabytes of data yearly from particle collision experiments and simulations. Exabytes of data are required to be processed, including metadata, and data from a posteriori analysis. Therefore, a huge amount of computing resources is needed for data storage, and to support a computing throughput of around 105 tasks per day followed by an increasing demand for efficient data sharing among computing centers through high-speed networks.
The Worldwide LHC Computing Grid (WLCG) has been created to support HEP experiments at CERN. The grid infrastructure is an essential asset to support the LHC discoveries. Nonetheless, grid resource requests tend to boom in the near future due to a scheduled LHC upgrade that aims to increase the experiment's luminosity by a factor of 10 over its current value, increasing the amount of data to process. In HEP scattering, luminosity 
is the ratio of events 
detected through a cross-section 
over a period of time 
, i.e., 
. The increase of 
generates at least a linear or polynomial increase on the amount of data, and, consequently, a polynomial increase of the workload of computing centers across the grid. A complex technological challenge is envisioned, namely, to keep the grid infrastructure working along the Run-3 and Run-4 stages of the High-Luminosity LHC project (HL-LHC) (Di Girolamo et al., 2020; Herr & Muratori, 2006).
The HEP Software Foundation (HSF) released a road-map document describing the actions needed to prepare the grid to support the HL-LHC upgrade (Albrecht et al., 2019). As a result, the Operational Intelligence group (OpInt) was created as a task force to improve the WLCG quality of service (QoS). Through data analytics and log data mining, its main research line concerns the development and maturation of machine learning (ML) tools based on event-oriented maintenance systems. Many ad-hoc solutions have been promoted by the OpInt group, from log parsing to diagnostic systems, as real-time anomaly detection approaches developed to assist the computing center of the Italian Institute of Nuclear Physics (INFN-CNAF) (de Sousa et al., 2019). ML algorithms reduce system downtime and optimize the usage of resources.