[CS] Ray Sinurat Candidacy Exam/Jan 6, 2026

[via cs]

This is an announcement of Ray Sinurat's Candidacy Exam.
===============================================
Candidate: Ray Sinurat

Date: Tuesday, January 06, 2026

Time:  1 pm CST

Remote Location: https://uchicago.zoom.us/j/99323041659?pwd=LjhLY4NTLekiemr4ojez1tEvWrT4F3.1  Meeting ID: 993 2304 1659 Passcode: 475923

Location: JCL 298

Title: Taming I/O Optimization for Deep Learning at Scale

Abstract: Scientific deep learning at scale typically trains on terabyte-scale datasets across thousands of accelerators, placing immense pressureon storage system to keep pace with computation. Existing solutions respond to this demand by tuning individual I/O pipeline parameters to accelerate training performance. However, these techniques are limited by pigeonholing tuning knobs in isolation, configuration space explosion, and expensive tuning, leading to suboptimal configurations that sacrifice training efficiency, system utilization, or both. 

To this end, we developed SysX, a cross-layer I/O optimization framework that jointly optimizes across application and system layers while keeping exploration cost tractable. SysX achieves this through three novel features: a unified metric that balances application throughput with system utilization, optimizer that efficiently narrows configuration space, and approximation-based technique that estimates full-pipeline behavior at a fraction of the runtime cost. We evaluate SysX on five weather forecasting workloads and demonstrate up to 2x improvement in application throughput and I/O bandwidth, 19x reduction in configuration space, optimization time cut from months to days, and order-of-magnitude reduction in storage footprint during optimization, all without requiring prior data collection. Overall, SysX delivers 1.3x to 1.8x faster training compared to original configurations on large-scale HPC system.


Advisors: Haryadi Gunawi

Committee Members: Haryadi Gunawi, Kyle Chard, Hariharan Devarajan, and Philip Carns