Power Management Techniques for Data Centers: A Survey

With growing use of internet and exponential growth in amount of data to be stored and processed (known as ‘big data’), the size of data centers has greatly increased. This, however, has resulted in significant increase in the power consumption of the data centers. For this reason, managing power consumption of data centers has become essential. In this paper, we highlight the need of achieving energy efficiency in data centers and survey several recent architectural techniques designed for power management of data centers. We also present a classification of these techniques based on their characteristics. This paper aims to provide insights into the techniques for improving energy efficiency of data centers and encourage the designers to invent novel solutions for managing the large power dissipation of data centers.

💡 Research Summary

The paper addresses the escalating power consumption of modern data centers, driven by the explosive growth of internet services, big‑data workloads, and cloud computing. It begins by quantifying the problem: in 2006 U.S. data centers and servers consumed 61 billion kWh (≈1.5 % of national electricity) costing $4.5 billion, and today many facilities host tens of thousands of servers drawing tens of megawatts. The authors identify the primary causes of high power draw: low average server utilization (10‑50 %), high power‑density leading to costly cooling (0.5‑1 W of cooling per watt of IT power), and the resulting reliability and environmental concerns.

The survey classifies power‑management techniques into four broad categories:

1. Dynamic Voltage and Frequency Scaling (DVFS).
   DVFS exploits the quadratic relationship between supply voltage and dynamic power (P ≈ C·F·V²). The paper reviews several adaptive algorithms that adjust voltage/frequency in response to workload intensity while respecting QoS constraints. Notable works include Sharma et al.’s Linux‑kernel feedback loop for web servers, Hsu & Feng’s bounded‑performance‑loss model, Horvath et al.’s coordinated multi‑tier voltage scaling, and Deng et al.’s extension of DVFS to DRAM channels and memory controllers. The authors note that DVFS is most effective for memory‑bound workloads but can degrade performance if applied indiscriminately.

2. Server Consolidation and Power‑State Transition.
   Because many servers run at low utilization, consolidating workloads onto fewer machines and powering down idle nodes can yield large savings. The survey discusses heterogeneous platform consolidation (Chun et al.), ensemble‑level power sharing (Ranganathan et al.), “barely‑alive” servers that keep only memory active (Anagnostopoulou et al.), and out‑of‑band management processors used to offload I/O (Ghosh et al.). Techniques also include live VM migration (Liu et al.) and per‑core power gating (Leverich et al.), often combined with DVFS for additive benefits. Limitations include transition latency, potential SLA violations, and the overhead of migration.

3. Workload Scheduling and Task Allocation.
   Scheduling approaches monitor incoming request rates and predict future demand to dynamically allocate servers. Chase et al. propose a reconfigurable switch that routes traffic to a selected subset of servers, balancing cost and service quality. Rusu et al. present a cluster‑wide QoS‑aware algorithm that accounts for boot‑up time when deciding which nodes to power on/off. These methods aim to flatten peak power, reduce unnecessary provisioning, and maintain SLA compliance.

4. Thermal‑Aware and Cooling Management.
   Thermal‑aware techniques use temperature sensors and heat‑distribution models to relocate workloads, adjust fan speeds, or control chilled‑water flow, thereby reducing cooling power. Some works explore integration with renewable energy sources and energy‑storage systems, while others target specific subsystems such as disks or main memory.

The authors critically evaluate each class, highlighting trade‑offs: DVFS can impair performance; server power‑state transitions incur latency and wear; scheduling relies on accurate workload forecasts; thermal methods depend on sensor fidelity and model accuracy. Moreover, most studies are validated via simulation or small‑scale testbeds, leaving open questions about scalability and interaction effects in real‑world hyperscale facilities.

In the concluding section, the paper outlines future research directions: (i) unified frameworks that jointly optimize energy, performance, temperature, and cost; (ii) machine‑learning‑driven predictive control for real‑time adaptation; (iii) deeper integration of renewable energy and on‑site storage; and (iv) development of standardized benchmarks and publicly available datasets to enable reproducible, large‑scale evaluation.

Overall, the survey provides a comprehensive taxonomy of architectural‑level power‑management strategies, summarizes key contributions and limitations of existing work, and offers a roadmap for advancing energy‑efficient data‑center design.