An opportunity to improve Data Center Efficiency: Optimizing the Server's Upgrade Cycle
Panagiota Nikolaou, Freddy Gabbay, Jawad Haj-Yahya, Yiannakis Sazeides
TL;DR
This work tackles data-center efficiency by optimizing the server upgrade cycle to maximize the metric $QPS/(TCO × CO2)$. It adopts a global, design-time upgrade plan that leverages both current and anticipated future servers and permits variable-length upgrade intervals, contrasting with local fixed-cycle strategies. An exhaustive planner applied to 11 servers (2010–2022) using the ACT CO$_2$ model demonstrates a ~19% gain in the target metric over the best local plan, with a four-server schedule (A:1, B:2, D:4, H:5) achieving optimal results. The findings underscore the value of anticipating future hardware in upgrade planning and point to scalable optimization, predictive hardware modeling, and more realistic lifetime emissions as pathways toward practical deployment in hyperscale data centers.
Abstract
This work aims to improve a data center's efficiency by optimizing the server upgrade plan: determine the optimal timing for replacing old servers with new ones. The opportunity presented by this approach is demonstrated through a study based on historical server data. The study establishes a significant opportunity to increase the QPS/(TCOxCO2) metric by formulating a global upgrade plan at the data center's design time covering its entire life cycle. This plan leverages information, such as server entry year, performance, and active power consumption for both existing and future servers. Our findings reveal that an optimal global upgrade plan, may involve upgrades at non fixed time periods and outperforms local upgrade plans. Local upgrade plans follow a fixed, equal-length cycle and make decisions based only on currently available server models. These local plans select the best available server at each upgrade cycle without accounting for future server releases.