000001858 001__ 1858
000001858 005__ 20240523045548.0
000001858 0247_ $$2doi$$a10.6082/uchicago.1858
000001858 041__ $$aen
000001858 245__ $$aMaximizing Performance in Power-Constrained Computing Systems
000001858 260__ $$bThe University of Chicago
000001858 269__ $$a2019-06
000001858 300__ $$a125
000001858 336__ $$aDissertation
000001858 502__ $$bPh.D.
000001858 520__ $$aPower constraint has become arguably the biggest obstacle for the performance scaling of computing machines. No matter what scale of computing system is – A mobile phone or supercomputer – they are all power restricted in one way or another to ensure normal operation. While various computing systems may require different power management technique, the goal of such systems is invariant and contains two folds of requirement: (1) guarantee computing system operating under a certain power budget/cap, and (2) make use of the limited power efficiently to deliver high performance. Thus, the challenge can be formalized to a classic constrained optimization problem – Given power consumption constraints, maximize the performance of computing systems. In this dissertation, we focus on solving this problem for server systems from single-node level to large-scale. More specifically, this dissertation contains 3 projects addressing power capping challenge at different spectrum.

First, we propose PUPiL, a hardware/software hybrid power control system to address the power challenge at the node level. It makes the key observations of tradeoffs between existing software-based and hardware-based approach:(1) hardware techniques provide significantly faster response time – quickly enforcing power limits and, (2) software provide much greater flexibility – by tailoring resource usage to the current application workload – leading to high performance efficiency. PUPiL combines the best of software and hard- ware approach, achieves significantly higher performance with nearly same response time as hardware approach.

Second, we propose PowerShift, a distributed power management system to address the emerging challenge of power capping dependent applications in large-scale system. Pow- erShift, to our knowledge, is the first work to identify the unique challenge of dependent distributed workloads and presents a family of three techniques for this scenario, demonstrating improved performance, reduced energy, and dynamic adjustment to tail behavior and system noise.

Last, PoDD, a hierarchical distributed power control system inspired by both PUPiL and PowerShift, is proposed to further overcome major limitations in power capping dependent applications. It incorporates learning/hardware hybrid node-level power capping with system-level power shifting to deliver significantly higher performance than prior works and no longer requires offline application profiles by build power model online, greatly improving practicality and performance efficiency.

The 3 power management framework systematically studied the problem of maximizing performance in power constrained systems. The key ideas and insights are highly general to guide design of real world power control system for wide range of workloads and platform. All implemented systems are open-sourced and evaluated to be practical, scalable, reliable and also not limited to particular applications and systems, which hopefully will serve as a base model/system to future research on power capping.
000001858 542__ $$fCC BY
000001858 650__ $$aComputer science
000001858 653__ $$aAdaptive System
000001858 653__ $$aClassification
000001858 653__ $$aPerformance Model
000001858 653__ $$aPower Management
000001858 653__ $$aResource Management
000001858 690__ $$aPhysical Sciences Division
000001858 691__ $$aComputer Science
000001858 7001_ $$aZhang, Huazhe$$uUniversity of Chicago
000001858 72012 $$aHenry Hoffmann
000001858 72014 $$aIan Foster
000001858 72014 $$aHaryadi Gunawi
000001858 8564_ $$9980c6efa-14e3-4806-86a5-c08f05e75c9d$$ePublic$$s1358229$$uhttps://knowledge.uchicago.edu/record/1858/files/Zhang_uchicago_0330D_14801.pdf
000001858 909CO $$ooai:uchicago.tind.io:1858$$pDissertations$$pGLOBAL_SET
000001858 983__ $$aDissertation