Speaker
Description
The ATLAS experiment at the CERN Large Hadron Collider (LHC) records and processes large amounts of data from proton-proton collisions. With the upcoming High-Luminosity LHC (HL-LHC), the data volume is expected to increase by more than an order of magnitude, posing new challenges for storage, data throughput, and analysis scalability.
To meet this challenge, ATLAS is transitioning to RNTuple, the next-generation data storage architecture designed to replace the legacy TTree. Currently, all major production output formats support RNTuple. Performance studies using ATLAS data have already demonstrated substantial improvements in space usage and I/O performance.
This study presents a comprehensive performance evaluation of the ROOT TTree and RNtuple storage layers, integrated with modern compression suites including LZMA, ZSTD, LZ4, and ZLIB. By benchmarking multi-processing scaling, throughput, and memory footprints (MVEM/RSS/PSS) within the ATLAS Athena framework, we characterize the operational trade-offs of current data formats.
However, as traditional high-dimensional data representations become unsustainable under future trigger and storage constraints, we explore integrating Machine Learning (ML) to move beyond classical Algorithms listed above. We then discuss early results from the transition toward ML-based dimensionality reduction using autoencoders and Mamba networks for data compression.
We investigate the hypothesis that ML-based compression algorithms can serve as anomaly detection, since when compression fails, it may mean that the data was not present in their training. We will test this hypothesis using new physics models that include a new force similar to the strong force, leading to new signatures and potential discoveries.
| Apply for student award at which level: | PhD |
|---|---|
| Consent on use of personal information: Abstract Submission | Yes, I ACCEPT |