Event-Driven Application Brownout: Reconciling High Utilization and Low Tail Response Times

Abstract

Abstract-Data centers currently waste a lot of energy, due to lack of energy proportionality and low resource utilization, the latter currently being necessary to ensure application responsiveness. To address the second concern we propose a novel application-level technique that we call event-driven Brownout. For each request, i.e., in an event-driven manner, the application can execute some optional code that is not required for correct operation but desirable for user experience, and does so only if the number of pending client requests is below a given threshold. We propose several autonomic algorithms, based on control theory and machine learning, to automatically tune this threshold based on measured application 95th percentile response times. We evaluate our approach using the RUBiS benchmark which shows a 11-fold improvement in maintaining response-time close to a set-point at high utilization compared to competing approaches. Our contribution is opening the path to more energy efficient data-centers, by allowing applications to keep response times close to a set-point even at high resource utilization.

BrownoutCC: Cascaded Control for Bounding the Response Times of Cloud Applications

“…The contribution of this paper is three-fold: (i) We formalize the brownout control strategy in [8] into a cascaded control problem; (ii) We design the inner and outer loop controllers, proposing both a feedback and a feedforward plus feedback version. Our controller features both the performance of [8] and the formal guarantees of [19]; (iii) We evaluate our approach and compare with previous solutions using the brownout simulator.…”

“…Our controller features both the performance of [8] and the formal guarantees of [19]; (iii) We evaluate our approach and compare with previous solutions using the brownout simulator. Besides providing formal guarantees, our controllers show fewer oscillations and maintain the measured response times closer to the target.…”

“…THE BROWNOUT APPROACH This section provides some background information about the brownout model and controllers developed in [8,19]. It also introduces some basic terminology that will then be used to explain the Brownout CC approach.…”

“…The event-based version of the brownout paradigm [8] works as follows. A periodic controller updates a threshold value ψ q (k) for the length of the queue of requests that have not yet been answered, with period τ c .…”

“…This could mean deciding based on the average of many response times or of none. To avoid this disparity and gain additional performance, an event-based version of the brownout principle was then devised [8], which would take a new decision at every request arrival. The event-based brownout controller [8] showed very good performance, but lacked the mathematical formalization and analysis possibility.…”