The EPOC Project - Energy Proportional and Opportunistic Computing System

Nicolas Beldiceanu, Barbara Dumas Feris, Philippe Gravey, Md Sabbir Hasan, Claude Jard, Thomas Ledoux, Yunbo Li, Didier Lime, Gilles Madi-Wamba, Jean-Marc Menaud, Pascal Morel, Michel Morvan, Marie-Laure Moulinard, Anne-Cécile Orgerie, Jean-Louis Pazat, Olivier H. Roux, Ammar SharaihaSMARTGREENS 2015 · DOI

TL;DR

This paper introduces the EPOC project, a system designed to optimize energy consumption in small to medium-sized data centers by intelligently integrating renewable energy sources. It proposes a novel approach to resource management that dynamically schedules tasks based on energy availability, aiming to reduce operational costs and environmental impact without sacrificing service quality.

What Problem Does This Paper Address?

The ever-increasing demand for online services has led to a dramatic rise in data center energy consumption. This poses significant challenges: environmental concerns due to CO2 emissions, strain on electrical grids, and escalating operational expenditures. While some companies are exploring renewable energy sources, their intermittent nature complicates consistent power supply. Existing solutions often fail to consider electricity availability and renewable energy integration, leading to energy losses during transmission and storage. This paper tackles the problem of efficiently powering data centers by leveraging both the regular electrical grid and local renewable energy sources, proposing a system that adapts to energy availability and optimizes task scheduling accordingly.

Methods

The EPOC project proposes a multi-faceted approach:

  • Energy Proportional Computing (EPC): The core idea is to design a system where energy consumption is directly proportional to activity, ideally reaching zero when idle. This is achieved through dynamic server on/off policies (Vary-On/Vary-Off or VOVO).
  • Virtualization and Live Migration: To maintain continuous service availability for online applications while enabling VOVO, the system utilizes virtualization and live migration. This allows for the dynamic movement of virtual machines (VMs) between physical servers, enabling servers to be powered down when not in use.
  • Hybrid Energy Approach: EPOC integrates both the regular electrical grid and local renewable energy sources (like solar or wind). It aims to utilize renewable energy whenever available, supplementing with grid power as needed.
  • Energy-Aware Task Scheduling: The system distinguishes between "web jobs" (requiring continuous operation) and "batch jobs" (which can be delayed or interrupted). Batch jobs are prime candidates for opportunistic scheduling during periods of high renewable energy availability.
  • Energy-Aware SLA Management: The project proposes an energy-aware Service Level Agreement (SLA) management system that considers energy as a first-class resource. This includes the concept of "virtual green energy," where surplus renewable energy can be used to offset periods of low availability, creating a perception of consistent supply for users.
  • High-Throughput Optical Networks: To support the massive data transfers required for VM migration (potentially terabytes), EPOC investigates energy-efficient optical network architectures. They propose the POPI (Passive Optical Pod Interconnect) architecture, which aims to reduce power consumption compared to traditional electrical packet switches.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • "Virtual Green Energy": This is a key conceptual innovation. Instead of relying on physical energy storage (which is costly and inefficient for small data centers), EPOC proposes to "virtualize" green energy. By intelligently scheduling tasks to utilize surplus renewable energy and then drawing from the grid when needed, the system can present a consistent energy supply to users, effectively masking the intermittency of renewables.
  • Opportunistic Task Scheduling: The paper highlights the significant potential of scheduling non-critical "batch jobs" during periods of abundant renewable energy. This allows for better utilization of green energy and reduces reliance on the grid during peak demand or when renewables are scarce.
  • The Necessity of High-Bandwidth Networks for VM Migration: The paper strongly emphasizes that migrating large VMs (especially their storage) requires extremely high network bandwidth. This drives the investigation into advanced optical networking solutions like POPI.
  • Storage as a Major Energy Consumer: The paper points out that storage devices account for a significant portion of a data center's energy consumption, underscoring the importance of efficient storage management and network architectures that can handle large data transfers without dedicated storage area networks (SANs).

Key Results

The paper presents the architectural principles and proposed solutions rather than specific quantitative results from a deployed system. However, it outlines expected benefits:

  • Reduced Operational Expenditure (OPEX): By leveraging renewable energy and optimizing task scheduling, the project aims to significantly lower electricity costs.
  • Lower Carbon Footprint: Increased use of renewable energy directly contributes to reducing CO2 emissions.
  • Efficient VM Migration: The proposed POPI optical network architecture is estimated to consume around 25% of the power of classical electrical packet switch architectures for interconnections, while supporting the high bandwidth needed for VM migration.
  • Improved Server Utilization: Dynamic consolidation and VOVO policies aim to increase server utilization and reduce idle power consumption.

Limitations and Future Work

  • Reliable Prediction Models: The success of opportunistic scheduling relies heavily on accurate prediction models for renewable energy availability. The paper acknowledges that developing such reliable models is crucial.
  • VM Migration Costs: While live migration enables energy savings, it incurs costs in terms of time and energy. The algorithms need to carefully balance these costs against the energy savings.
  • Scalability of POPI: The paper discusses the POPI architecture and its potential for small to medium data centers, with an estimate of 61 servers per length. Further investigation into its scalability for larger deployments might be needed.
  • Prototype Validation: The paper presents the EPOC project and its architectural design. The next logical step would be the implementation and validation of the EpoCloud prototype to demonstrate the effectiveness of these principles in real-world scenarios.
  • SLA Definition and Enforcement: While an energy-aware SLA is proposed, the specifics of defining and enforcing such SLAs, especially with the concept of "virtual green energy," require further development and practical testing.

Summary updated

BibTeX
@inproceedings{beldiceanu2015epoc,
  author = {Nicolas Beldiceanu and Barbara Dumas Feris and Philippe Gravey and Md Sabbir Hasan and Claude Jard and Thomas Ledoux and Yunbo Li and Didier Lime and Gilles Madi-Wamba and Jean-Marc Menaud and Pascal Morel and Michel Morvan and Marie-Laure Moulinard and Anne-Cécile Orgerie and Jean-Louis Pazat and Olivier H. Roux and Ammar Sharaiha},
  title = {{The EPOC Project - Energy Proportional and Opportunistic Computing System}},
  booktitle = {SMARTGREENS},
  year = {2015},
  doi = {10.5220/0005487403880394},
}
Figure