On The Fundamental Energy Trade-offs of Geographical Load Balancing

Abbas Kiani, Nirwan Ansari

TL;DR

This paper investigates the fundamental trade-offs between energy consumption, green energy utilization, and electricity costs in geographically distributed data centers. It proposes a novel information flow graph model to capture these trade-offs, suggesting that simply maximizing green energy or minimizing cost might not always lead to the lowest total energy consumption.

What Problem Does This Paper Address?

The ever-increasing demand for online services has led to a massive growth in data center energy consumption, contributing significantly to electricity bills and carbon emissions. While geographical load balancing (GLB) offers a promising approach to leverage diverse energy sources (renewable and grid) and electricity prices across dispersed data centers, existing strategies often focus on optimizing individual objectives like maximizing green energy use or minimizing electricity costs. However, the paper highlights that these isolated optimizations can lead to increased total energy consumption due to factors like varying server capabilities and network delays between data centers. This creates a need for a more holistic approach to understand and manage the complex interplay of these factors.

Methods

The authors introduce an Information Flow Graph (IFG) model to represent geographically dispersed data centers. This model uses directed acyclic graphs with three types of nodes: a source node (Front-End server, FE), intermediate nodes (data centers), and a data collector node (users). The core of their approach lies in defining a service efficiency parameter for each data center. This parameter is derived from an M/GI/1 Processor Sharing (PS) queue analysis, which accounts for network delay between the FE and data centers, as well as queuing and processing delays within the data center. This allows for a more nuanced evaluation of how efficiently a data center can handle requests, considering both its processing power and network latency. The model then aims to optimize workload distribution to balance total power consumption, green energy utilization, and cost.

Datasets

Not discussed in the paper. The paper uses simulated data based on trends of wind power and total workload from the 1998 World Cup website.

Key Insights and Ideas

  • The "Green vs. Brown" Dichotomy: The paper advocates for distinguishing between workloads served by green energy ("green workload") and those served by grid energy ("brown workload"). This separation is crucial for simultaneously maximizing green energy utilization and minimizing brown energy costs.
  • Trade-offs are Fundamental: Simply chasing more green energy or cheaper electricity isn't always the optimal path. Sending a request to a data center with higher network delay or lower service capability, even if it offers cheaper power or more green energy, can paradoxically increase overall energy consumption.
  • Service Efficiency is Key: A well-defined service efficiency parameter, incorporating network delays and queuing dynamics, is essential for making informed workload distribution decisions in a GLB context.
  • Information Flow Graph as a Unified Model: The IFG provides a powerful framework to capture the complex interactions and trade-offs between various objectives in a geo-dispersed data center environment.

Key Results

The paper presents simulation results demonstrating the trade-off curves between total power consumption and green energy utilization at different times of the day. These curves visually confirm that increasing green energy utilization generally leads to an increase in total power consumption. For instance, at 10 AM, utilizing all available green energy at 6 data centers resulted in over 35 Megawatts of total power consumption. The simulations also illustrate how workload allocation shifts based on wind power availability and electricity prices, showcasing the dynamic nature of optimal GLB strategies.

Limitations and Future Work

The authors acknowledge that their model currently considers only one FE server, though they note it can be extended to multiple FEs. Future work could explore:

  • Extending the model to handle multiple workload distribution centers (FEs).
  • Incorporating the availability and impact of energy storage devices (e.g., batteries) at data centers, including battery-related constraints.
  • Further elaborating on machine learning-based energy prediction techniques for more accurate forecasting.
  • Investigating the impact of battery-related constraints on energy storage.

Summary updated

BibTeX
@article{kiani2017fundamental,
  author = {Abbas Kiani and Nirwan Ansari},
  title = {{On The Fundamental Energy Trade-offs of Geographical Load Balancing}},
  journal = {IEEE Communications Magazine},
  year = {2017},
  doi = {10.1109/mcom.2017.1600787},
}
Figure