Learned Performance Model for SSD

Abstract

The advent of new SSDs with ultra-low latency makes the validation of their firmware critical in the development process. However, existing SSD simulators do not sufficiently achieve high accuracy in their performance estimations for their firmware. In this paper, we present an accurate and data-driven performance model that builds a cross-platform relationship between the simulator and target platform. We directly execute the firmware on both platforms, collect its related performance profiles, and construct a performance model that infers the firmware's performance on the target platform using performance events from the simulation. We explore both a linear regression model and a deep neural network model, and our cross-validation shows that our model achieves a percent error of 3.1%, significantly lower than 18.9% from a state-of-the-art simulator. © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.