Georgia Institute of Technology
Maximizing Datacenter Power Efficiency Through Local Load Imbalancing
Abstract
dc:description.abstractModern datacenters typically distribute incoming traffic evenly across over-provisioned servers to strictly meet latency Service Level Objectives (SLOs). However, because modern CPUs are not natively power-proportional, this balanced approach forces entire clusters to constantly pay a high baseline idle power tax. This thesis demonstrates that CPU power consumption actually scales sublinearly with load, meaning the marginal energy cost of processing requests decreases as server utilization increases. To exploit this hardware behavior, we propose the Slack-Aware Load Imbalancing Protocol, a distributed routing architecture that intentionally coalesces traffic onto a dynamically sized subset of active servers to drive them toward peak power efficiency. By safely maintaining remaining slack capacity in a low-power idle state and utilizing localized request bouncing to handle micro-bursts, our real-world deployment on a live cluster achieves up to 7% total power savings without violating strict latency SLOs.
Degree
thesis:*- Name thesis:degree_name
- Computer Science, MS
- Department thesis:degree_department
- Computer Science
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Thummar, Dhyey
- Advisor dc:contributor.advisor
-
- Saeed, Ahmed
- Committee members dc:contributor.committeemember
-
- Dhekne, Ashutosh
- Xu, Jun
Rights
- Language dc:language.iso
- English
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/81630