[Online]Recognizing Shared Operational Events Across Data Center Racks from Distributed Telemetry

Recognizing Shared Operational Events Across Data Center Racks from Distributed Telemetry
ID:87 Submission ID:64 View Protection:ATTENDEE Updated Time:2026-10-04 23:37:08 Hits:162 Online

Start Time:2026-10-12 16:15 (Asia/Ho_Chi_Minh)

Duration:15min

Session:[S5] Track 5: Emerging Trends of AI/ML » [S5-5] Track 5: Emerging Trends of AI/ML

Presentation File

Tips: The file permissions under this presentation are only for participants. You have not logged in yet and cannot view it temporarily.

Abstract
Modern data centers generate large volumes of component level telemetry, while operational decisions are made at the level of higher-level events. This paper presents a telemetry driven framework for recognizing shared operational events across distinct physical compute racks using publicly available production telemetry from the Marconi100 Tier-0 HPC system. Normal-to-non-normal node-state transitions are aggregated within racks, correlated across racks, and consolidated into operational episodes. Cross-rack synchronization is statistically evaluated using an exact rack-preserving whole-week alignment test, while telemetry trajectories characterize event behavior around onset. Across 45 nodes in three racks, 5,024 valid anomaly onsets produced 77 synchronized cross-rack timestamps and 20 operational episodes. Among the 15,375 alternative relative whole-week rack alignments, the maximum synchronization count was two, compared with 77 in the observed alignment ($p_{\mathrm{exact}}=1/15376\approx6.5\times10^{-5}$), and the episode count remained stable under alternative definitions. Trajectory analysis identified state-dominant or mixed, controlled-transition-like, and abrupt-loss-like modes. Post hoc comparison with CINECA records associated controlled-transition-like events with scheduled operational activity and one abrupt-loss-like event with a documented power-related disruption. Overall, the proposed framework aggregated distributed node-level monitoring anomalies into robust and interpretable shared operational events, providing a higher-level representation for infrastructure monitoring and Artificial Intelligence for IT Operations (AIOps).
Keywords
AIOps, anomaly correlation, complex event recognition, data centers, distributed telemetry.
Speaker
Ilyas Saleem
Technical Program Ma Amazon Web Services

Submission Author
Ilyas Saleem Amazon Web Services
Comment submit
Verification code Change another
All comments

CONTACT US

Email: asiancomnet2026@usssociety.org

Website & IT Support: hi@aconf.org

Registration Submit Paper