A Simulation-based Study on the Reliability of Data Center Upgrades

Abstract

Reliability is a critical concern for data centers and has attracted extensive studies. Prior studies mostly focused on static data centers, assuming that the software and hardware components remain unchanged for a long time. However, today's data centers have to upgrade due to the increasing requirements of security, performance, or capacity, and it remains unclear how the data center upgrades affect data reliability. In this paper, we particularly focus on studying typical upgrading policies that are widely adopted in data centers and present a simulation-based study to reveal the effects of different policies on data reliability. We first design and implement an event-based reliability simulator called SIMDDC, which simulates different system organizations and failure/recovery models, and takes the probability of data loss ( PDL) and the probability of data unavailability ( PUA) as the metrics of reliability. By using SIMDDC, we then conduct a set of experiments. Based on our quantitative analysis of the experimental results, our study has revealed several interesting and important findings related to software and hardware upgrades in terms of upgrade domain settings, upgrade checks, and upgrade with different storage devices (HDDs and SSDs) and different data redundancy schemes. Our findings are helpful to determine proper upgrade policies while guaranteeing the data reliability in modern data centers. For public use, we have open sourced our simulator SMIDDC at http://github.com/yichuan0707/SIMDDC.