[CS] Jianru Ding Candidacy Exam/Apr 22, 2025

[via cs]

This is an announcement of Jianru Ding's Candidacy Exam.
===============================================
Candidate: Jianru Ding

Date: Tuesday, April 22, 2025

Time: 10 am CST

Remote Location: https://uchicago.zoom.us/j/95567178640?pwd=f3ZkCHbfwDeEq7PgVvN7BdztVa62GD.1  
Meeting ID: 955 6717 8640 Passcode: 344933

Location: JCL 298

Title: Adaptive Movement of Schedulable Objects in Distributed Systems

Abstract: With the rapid increase of distributed workloads (e.g. emerging LLM workloads), and extensive expansion of distributed systems (e.g. the exa-scale clusters), scheduling is the key to efficiently harnessing the limited computing power of distributed systems. The thesis covers three aspects in scheduling: power, computation, and cooling. 

The constraints on power and energy consumption require resource allocation strategies to respect a cluster-wise power budget in an over-provisioned system. We showcase a low-overhead power scheduling system that improves the performance of distributed workloads with no reliance on a learned model as the first part of this work.

To schedule computation with heterogeneous data sources, we present an abstract work unit, the Unified-Work, to provide a homogeneous view of assignable tasks. Based on such abstraction, we introduce the load-balancing system that simultaneously schedules in-memory and streaming tasks, and extend the work-stealing scheme to a million-fold scale.

The final piece of the thesis is cooling scheduling. Cooling has become crucial in maintaining operational efficiency in the emerging power-hungry GPU data centers as the LLM workloads become prevalent. With scheduling the cooling intensity of liquid cooling units in data centers becoming an option in production, we propose to understand the cost-efficiency trade-off in cooling scheduling, and explore approaches that adaptively co-schedule the three schedulable objects correspondingly.


Advisor: Hank Hoffmann

Committee: Hank Hoffmann, Haryadi Gunawi, Yi Ding, Chaojie Zhang