Skip to main content

Green Active Management of Data Centers and Energy Aware Storage Management

The recent GAMES EU project aimed to develop an innovative framework of methodologies, tools and services, as well as energy efficiency assessment metrics, for the holistic energy-aware design and operations of next generation IT Service Centers. It developed methodologies, models and tools for green data centers starting from the early design phase of business processes, system architecture and applications, and continuing on to the runtime control of all aspects of data center computing. The energy optimizations of GAMES include both global control (for example, of a given service or set of services) as well as several local control loops (that control, for example, server behavior and the data center storage system). Hence, it considered all levels from applications to hardware/software and facilities such as cooling systems.

GAMES defined a new category of Key Performance Indicators (KPIs), named Green Performance Indicators (GPIs), aimed at providing energy efficiency and emissions metrics. GPIs are based on Green Metrics, namely measures enabling the analysis of energy losses in applications and the execution of actions that can save energy and involve the CPU, memory, I/O channels and storage. Once energy leakage is discovered (via monitoring) actions are triggered to reduce this energy loss by appropriate actions. Examples of actions could be reducing redundancies of data and processes, using storage in quiet mode, and changing the rate of CPU usage.

GAMES proposed a methodology for developing green applications that are annotated with KPI and GPI attributes. The local control loops (e.g., storage management system) may leverage these annotations to improve energy efficiency for the application.

GAMES includes a dedicated runtime monitoring subsystem and database to save the relevant data for immediate and further processing and data mining. At runtime, the servers' local controllers may perform adaptive actions such as re-configuring server components (e.g., processor voltage scaling and dynamic virtual machine allocation). The storage controller may change disk acoustic modes and control data placement in an energy aware manner. (See more details in SAMCEE below.)


Our major role in the GAMES EU project was the energy aware management of the storage subsystem (SAMCEE - Storage Advanced Management Controller for Energy Efficiency). The goal was energy-aware management for application storage. The major design decisions were

  • Separate control of data placement and disk acoustic modes,
  • File level data granularity with file-split into chunks,
  • On-line control of new file allocation and disk modes versus off-line data migration,
  • Spinning down of only unused disks while keeping reserve disks for performance, and
  • Data placement based on device ranking.

Device ranks were calculated based on our novel usage centric energy efficiency metrics for storage, which measure the energy efficiency of actual storage usage rather than the potential efficiency. Disk acoustic modes were controlled via fuzzy logic inference system based on insights from previous work.

SAMCEE includes two separate controllers: one for disk mode, and another for file placement and device sleep/start operation. Application level files are split into smaller chunks, and each chunk is placed separately on a selected device. The selection of the most appropriate location for chunk placement is based on device ranking, calculated according to the usage centric energy efficiency metrics for storage . Data consolidation and high device usage is attempted, allowing the switching off of unused devices (keeping spares for performance) and the saving of energy. SAMCEE includes a FUSE based file splitter

The evaluation of experimental results suggests that energy savings result from two scenarios: static and dynamic data access. For static data, migration that consolidates file chunks into a smaller number of active disks enables the spinning down of unused disks and save energy (0% – 25%). For dynamic data, depending on the data access patterns, energy savings are 0% –13%.

We concluded that:

  • Most of the energy savings is obtained by spinning down unused disks, which result from data consolidation and over-provisioning,
  • Employing the usage centric energy efficiency metrics for storage device ranking is beneficial, and
  • Data migration for data consolidation is beneficial only for the (quasi) static data scenario.

See SAMCEE presentation in attached file.