[Colloquium] Jianru Ding MS Presentation/Dec 7, 2022

[nitayack at cs.uchicago.edu]

This is an announcement of Jianru Ding's MS Presentation
===============================================
Candidate: Jianru Ding

Date: Wednesday, December 07, 2022

Time:  9 am CST

Remote Location: , https://uchicago.zoom.us/j/95126033378?pwd=NEpCQUt6TUtwbDYxZEtRaHJiblVMUT09 Passcode: 283773

M.S. Paper Title:  DPS: Adaptive Power Management for Overprovisioned Systems

Abstract: Distributed computing systems are subject to system-wide power limits, which are broken down into limits for each computing node. Constant power caps is a common practice. The runtime power consumption, however, is varying. Because of the different application being executed on each cluster and the different computing loads of each application phase, at a certain time, only a number of nodes are capped and the others are wasting the power budget allocated to them, which creates the problem of how to allocate the system-wide power budget so that the limit is respected and application performance is maximized. State of art works tackle this problem with either stateless systems or model-based approaches. Stateless systems do not need prior data to deploy, but they ignore the power changing speed and sequence of each node, and often fail to assure application performance due to lagging in cap adjustment and lacking ideas of node priorities. Model-based approaches improve application performance, but the need for a lot of hardware or application feedback data to build the models renders a huge deployment overhead. 

In this paper, we propose the Dynamic Power Scheduler (\name) that sustains the low-overhead advantage of model-free approaches and provides performance-boosting power cap allocations by analyzing the dynamics in a short online power history. Compared to constant power allocation, {\name} achieves at most 17.5\% increase in performance for compute-intense applications and guarantees no decrease for other applications under the same power budget.

Advisors: Hank Hoffmann