The Fifth Elephant 2025 Annual Conference (18 & 19 July)

As enterprises increasingly rely on Software-Defined Storage (SDS) to manage cloud-scale infrastructure, ensuring optimal workload placement, capacity planning, and performance monitoring has become a complex challenge. Traditional AI-driven observability solutions for IT infrastructure monitoring often rely on device-specific data and short-term statistics, which restricts their ability to generalize across workloads and storage environments. Recent advancements in Time Series Foundation Models (TSFM) present a promising alternative by leveraging pretrained representations on diverse long-term datasets, enabling transfer learning across devices and improving predictive accuracy.
Our work introduces a TSFM-based approach for AI-driven storage management, leveraging the TinyTimeMixer (TTM) architecture. Our solution, in production, incorporates domain-specific features to enhance TSFM pretraining for storage performance forecasting. Through extensive experiments on real-world FlashSystem performance metrics data, we demonstrate significant improvements over traditional machine learning baselines. These results highlight the potential of TSFM-powered observability to enhance automated storage management.

In this talk, we will dive into the steps taken to customize general purpose time series foundation models like TTM to solve storage domain performance observability use case in production.

Outline

• TTMs - Brief introduction
• Storage domain performance observability
• Extended-pretraining of TTMs
• Performance analysis compared to classical forecasting models
• Deployment status

Key Takeaways

• Learn about Time Series Foundation Models
• Customization of TSFM, in-particular TTM, for storage usecases
• Naunces of TTM in production…

Speaker Bio

Chandramouli Kamanchi is a Research Scientist working at IBM Research Bangalore, with over 6 years of industry experience.
He obtained his PhD from IISc Bangalore. His primary areas of interest are Reinforcement Learning, Stochastic Approximation, Time Series Forecasting. His expertise lies in solving real world problems through mathematical modeling. Beyond his professional pursuits, he is an amateur chess player, and enjoys cooking.

Profile links

[https://scholar.google.co.in/citations?user=1QlrvHkAAAAJ&hl=en]

Resources

TTM - Publication [https://proceedings.neurips.cc/paper_files/paper/2024/file/874a4d89f2d04b4bcf9a2c19545cf040-Paper-Conference.pdf]

TTM in Storage Insights [https://www.redbooks.ibm.com/redpieces/pdfs/redp5755.pdf]

Presentation

Slide deck is available here