Opportunities of Renewable Energy Powered DNN Inference

TL;DR

This paper explores how to manage deep neural network inference performance under fluctuating renewable energy supply by analyzing control mechanisms and proposing future research directions.

Contribution

It provides empirical insights into the impact of control knobs on DNN inference performance amidst renewable energy variability.

Findings

Different control knobs affect throughput variably under power fluctuations

Empirical profiling reveals key factors influencing inference performance

Trace-driven simulations guide future research directions

Abstract

With the proliferation of the adoption of renewable energy in powering data centers, addressing the challenges of such energy sources has attracted researchers from academia and industry. One of the challenging characteristics of data centers with renewable energy is the intrinsic power fluctuation. Fluctuation in renewable power supply inevitably requires adjusting applications' power consumption, which can lead to undesirable performance degradation. This paper investigates the possible control knobs to manage the power and performance of a popular cloud workload, i.e., deep neural network inference, under the fluctuating power supply. Through empirical profiling and trace-driven simulations, we observe the different impact levels associated with inference control knobs on throughput, under varying power supplies. Based on our observations, we provide a list of future research directions to leverage the control knobs to achieve high throughput.