A group of information scientists representing universities in the US, Australia, and China have created a novel task scheduling algorithm that improves cloud computing energy management. Engineered based on strategic decision-making principles inherent in game theory, the research paper illustrates how the proposed algorithm can “coordinate the distribution of tasks and the distribution of energy” better than current task scheduling methods.
As we’ll discuss, the cutting-edge algorithm achieves a substantial increase in task processing speed as compared to alternatives. The researchers say this speed boost coupled with eliminating computational bloat from unnecessary operations and scheduler communication may afford data center providers greater control of energy management.
In this article, we’ll discuss the global energy impact of data centers and provide a high-level overview of the proposed energy saving algorithm outlined in the paper.
Current State of Cloud Energy Usage
Measured as a whole, if the world’s IT sector were its own country, it would rank as the third-largest energy consumer on the planet. Think about that. Only the US and China – the world’s leading superpowers – would outrank this hypothetical nation in terms of annual energy usage.
It’ll come as no surprise that data centers make up a sizable portion of the total power consumed by the tech industry. VxChange – an American colocation and data center provider – says US-based data centers alone require north of 90 billion kilowatt-hours per year. This figure will likely balloon as operators build out resources to accommodate new products and services based on fast-rising technologies like big data, machine learning, IoT, and 5G.
Most would agree that steadfast technical achievement is a clear positive for society. On the other hand, however, the greenhouse gases released to generate the power needed to sustain our appetite for constant growth is a clear negative. Each internet search, video stream, and social media post requires energy, and, thus, is tethered to an amount of gas waste released into the atmosphere. Next time you use Google, consider that a typical search “requires as much energy as illuminating a 60-watt light bulb for 17 seconds and typically is responsible for emitting 0.2 grams of CO2.”
Mainstream adoption of cloud technology has sparked a surge in interest in the subject of energy management among operators. Advances in storage technology and machine virtualization, for example, have contributed greatly to data center infrastructure efficiency. In fact, NetworkWorld says “virtualization has meant a six-fold increase in compute instances with only a 25% increase in server energy use.” However, as policymakers continue to impose new energy regulations, and consumers increasingly demand that companies become more proactive regarding sustainability efforts, providers will need to develop an all encompassing approach to handling inefficiencies.
This is why efficiency gains through cloud-first hyper-optimization is growing in popularity among academics and data center professionals alike. Considering the vastness of modern data centers and the power of compounding, operators are discovering that even the slightest optimization can reap lucrative energy-savings.
Novel Game Theory Scheduling Algorithm Architecture Explained
To get a better understanding of the proposed algorithm and why it outperforms established techniques, we first have to take a look at its underlying architecture.
As mentioned, the proposed algorithm is rooted in game theory, a branch of mathematics focused on decision-making among interacting participants. According to the paper, “game theory is used to quantify the interaction between incentive structures, and to study the interaction between different players in strategic interdependence. Game theory considers the forecast behavior and the actual behavior of the participants in the process, and studies their optimization strategies”
A “player” is a task and a compute node – i.e. virtual machine or physical server – is the “game strategy.” Gameplay involves finding the optimal task schedule considering the available compute capacity (game strategy) and the preference, intent, and priority of each task (player).
With this mind, we can now deconstruct the algorithm to further examine its inner workings. The paper outlines eight distinct actions required to produce a hyper-optimized schedule, including (1) User request, (2) Parameter initialization, (3) Game process, (4) Potential function, (5) Nash equilibrium, (6) Conditional judgment, (7) Optimal strategy, and (8) Scheduling results.
Action 1: User request – The scheduling process begins when the user submits a task.
Action 2: Parameter initialization – A tree structure based on player and game strategy information output by Action 1 is generated and initialized.
Action 3: Game process – The game process is initialized but paused, awaiting an outcome “based on Nash equilibrium for the Potential function”.
Action 4: Potential function – Using a (separate) fixed function outlined in the paper, the algorithm calculates the lowest point of the Potential function. The lowest point is used to determine the stability of the game.
Action 5: Nash equilibrium – Nash equilibrium is used to optimize the Potential function. The purpose is to encourage cooperation among players.
Action 6: Conditional judgment – The algorithm determines if a set of requirements is met before moving forward. If requirements aren’t met, the algorithm will restart from Action 2.
Action 7: Optimal strategy – The next step is to optimize the schedule task span. This is defined as “the total time required by the system to complete the service requests submitted by the cloud user”.
Action 8: Scheduling results – Lastly, the solutions and calculations of select actions converge to output a hyper-optimized distribution of task states.
Research Results and Consequences of Cloud-First Engineering
When compared against ant colony and genetic task scheduling algorithms, the game-based algorithm proved superior. The researchers conducted a range of experiments, for instance, testing on compute nodes with high calculation ability, as well as nodes with balanced calculation ability and test varying the number of schedulers. Their findings show “no matter what kinds of computing power, the algorithm proposed in this paper can achieve the best optimization”.
Building a more efficient future will require data centers to do more work with less energy. Therefore, providers, in concert with academia, must continue to focus on uncovering and eliminating data center waste at every level of operation. This research demonstrates that measurable energy savings can result from rethinking established methods and techniques with a cloud-first engineering mindset.
##
About the Author
Daniel Browning is the Business Development Coordinator at PDF Supply, a global supplier of automation products. He enjoys writing about automation, AI, and new technology.
