Project - Topics
Within FIT4Green first ideas regarding green service level agreements (GreenSLA) are being developed that shall serve as steering instruments for the energy efficient workload management in a data centre.
SLA are the technical part of a contract between IT service provider and customer that defines the performance requirements of the service delivery as well as the service itself, the contractual partners as well as penalties. Usually the performance is tuned according to the needs of the customer, the cost for the provider, and the general practice of the particular IT service market. As generally energy efficient workload management conflicts with performance, the flexibility for a data centre to pursue energy efficiency goals through thei
Within FIT4Green first ideas regarding green service level agreements (GreenSLA) are being developed that shall serve as steering instruments for the energy efficient workload management in a data centre.
SLA are the technical part of a contract between IT service provider and customer that defines the performance requirements of the service delivery as well as the service itself, the contractual partners as well as penalties. Usually the performance is tuned according to the needs of the customer, the cost for the provider, and the general practice of the particular IT service market. As generally energy efficient workload management conflicts with performance, the flexibility for a data centre to pursue energy efficiency goals through their workload management is diminished.
In this context, GreenSLA are an approach to take nature into account and create a win-win-situation for all partners by
- relaxing the limits of performance oriented SLA metrics and thus extending the scope for energy saving measures in the data centre workload management
- introducing environmental KPI that on the other hand limit the contractual energy consumption of the job execution in question
- offering monetary and non-monetary incentives, tailored to specific services and user groups, that serve as balance between data centre customers’ and data centre management optimization goals (see figure 1).
The approach provides a win-win situation for all parties involved: The data-centre provider, the customer and the environment. Through GreenSLAs the data centre provider can save energy and thus save costs. These financial savings can, in turn, be shared with the customer while the energy savings reduce carbon dioxide emission and thus alleviate environmental pressure.
FIT4Green discusses options to model GreenSLA specifically for the data centre services of the computing infrastructures tested within the project: Traditional computing, cloud computing and supercomputing.
Further research will be directed towards
- Defining the trade-off between performance losses and energy savings
- Defining new energy metrics and saving strategies for SLAs
- Creating business models integrate GreenSLAs into their overall policies
- Implementing and testing of the GreenSLAs in their context
FIT4Green has baseline scenarios in the areas of cloud computing, traditional computing and supercomputing to show the setting in which the FIT4Green strategies should operate and thus serve as proof-of-concept (see figure 1). For all three computing styles, there is a single site and a federated site version.
- Cloud Computing Scenarios: SQ 3_1 (single site), SQ 3_2 (federation) at HPIS in Milano
- Supercomputing scenarios: SQ 2_1 (single site), SQ 2_2 (federation) at FZJ in Jülich
- Traditional scenarios: SQ 1_1 (single site), SQ 1_2 (federation) at ENI in Milano
In addition to the internal scenarios three external scenarios were created in order to give an impression about deployment options for FIT4Green in a real environment:
Accelerated IT:
A data centre in Germany with two different business models: offering web content services and social networking on the one hand and collocation services (infrastructure, dedicated servers, management, software as a service and consulting) on the other hand. Within the first business model, options to realize server consolidation, also based on long-term peak analyses are deemed realistic. For the second part of the business, security issues and skeptical customers make the realization of these strategies much more difficult. In case of a cluster of dedicated servers, however, it may well be possible, in collaboration with the data center customers, to create private clouds where consolidation is possible.
Luna.nl:
An Almende daughter-company based in the Netherlands that provides data centre services. The company lies on the border between traditional and cloud computing following the IaaS paradigm. Luna’s focus lies on Infrastructure Outsourcing, Managed Hosting, SaaS Partnerships, High Availability Solutions and Virtual Office. In addition to Luna’s direct customers, the company also offers services to customers of other Almende daughter-companies, such as ASK Community Systems, active in the emergency response service market, and Sense Observation Systems, active in the mobile computing and (application) service market. Having two data centres in Amsterdam and Rotterdam, the company sees opportunities to deploy FIT4Green federation strategies as well as workload consolidation on few or specially efficient servers and shutting down unused servers.
FIT4Green Dissemination Activities: Energy-Efficient Data Centres E2DC
FIT4Green project has initiated an annual international workshop in order to increase the awareness of green solutions for the data centres and the impact of energy-aware technologies for data centre business. The first International workshop on Energy-Efficient Data Centres (E2DC) will be held in May 2012 in conjunction with e-Energy conference.
The energy consumption of data centres offering all kinds of services such as cloud services, high performance computing, and traditional storage and data management in global organizations is increasing exponentially. For the sustainable development of ICT technology, new solutions for energy-efficient computing are needed in various technology fields. Especially the data centres play a big role in targeting energy-efficient and low-carbon economic development.
The E2DC aims to increase the international awareness of energy-aware data centre technologies and to allow the exchange of ideas and interaction among the different researchers in the topic. The scope of the workshop reaches from research on green data centre technologies varying from the information and communication technologies to business models and Green SLA solutions. The workshop also addresses researchers from infrastructure research to share their ideas and solutions for green data centres.
The main topics include for example:
- Energy-aware communication and network solutions for data centres
- Energy-aware cloud and high performance computing solutions
- Energy-aware data storages and reservoirs
- Scalability of federated data centres
- Large scale simulations
- Energy-aware data centre infrastructures and architectures
- Novel business models of green ICT and data centres
- Energy optimization algorithms and models
- Green SLAs
- Green computation and services
Power Consumption Prediction Models
In FIT4Green, energy savings can be achieved in data centres through optimization mechanisms whose main objective is to minimize the energy consumption as well as CO2 emissions. However, in order that these optimization mechanisms can take the most suitable energy-saving decisions, the existence of accurate power consumption prediction models becomes primordial. Consequently, the main focus of Work Package 3 is to provide such prediction models for ICT resources (e.g., servers, storage devices and network equipment) of data centres. The accuracy of the predictions of these models is crucial to take the most appropriate energy-saving decisions; in fact it has the responsibility to forecast the power consumption of a data centre after a possible reconfiguration option.
As a first step, a generic FIT4Green meta-model (schema) is provided where the most relevant energy-related ICT resources, their interconnections and the respective load applied on the infrastructure are represented. Given the complexity and heterogeneity of the data centre infrastructure, we introduced a methodology to derive the meta-model by decomposing the modelling process into 4 phases: Data centre Modelling, Server Modelling, Storage Modelling, and Services Modelling. Furthermore, the most relevant energy related attributes (both static and dynamic) of the resources have been identified in such a way that the monitoring system of data centres can provide reliable information. On one hand, static attributes are those that do not change throughout the execution period of the plug-in; for instance, the size and number of memory modules. On the other hand, a dynamic attribute is the one whose value changes frequently during the run-time and is usually based on the corresponding resource’s utilization (e.g., the load of a processor, the RPM of the fans, etc.).
Also, models were proposed for servers, storage devices and network equipment. Regarding servers, a component-based power consumption model is introduced where separate models for processors, memories, hard disks, fans and network interface cards were developed. To this end, both idle (no activity) and dynamic (utilization-based) aspects of power consumption were taken into account for the above-mentioned components of a server.
With respect to storage devices, a generic model is proposed that takes into consideration all types of storage devices from a single hard disk to Storage Area Network (SAN) devices. Since it’s challenging for the monitoring system of data centres to provide information such as when the hard disk changed from startup, to idle or accessing modes, then we adopt a probabilistic approach based on the current as well as maximum read and write rates.
Finally, a generic model is given for estimating the energy cost of transferring data over network equipment such as routers and switches
The Optimizer is a key component of the FIT4Green plug-in. It can be triggered for a Virtual Machine (VM) Single Allocation, in which case it will find the best allocation solution in terms of energy consumption, while respecting the SLA and QoS. It can also be triggered for a Global Optimization, where it will perform a complete data centre reconfiguration.
The engine of Optimizer is based on the Constraint Programming (CP) paradigm. The main features of CP are flexibility and scalability: indeed in CP the user’s requirements and operator’s requirements are expressed in their own Domain Specific Language (DSL) in the form of constraints. These constraints are clearly separated from the optimization algorithms and, therefore, are interchangeable and easily extendable. In order to derive a proper solution on Virtual Machine (VM) allocation and reallocation, Optimizer’s CP Engine processes the data from four inputs.
One of the inputs of the Optimizer is the current data centre configuration, which is derived from the current model instance, an extensive description in XML of the data centre components. Every time a request is made to the Optimizer, either a single allocation of a VM or a global optimization, the current model instance is passed as a parameter.
Power objective is constructed by using the energy consumption formulas. The CP engine needs an objective function to minimise; this is the so-called Power objective.
A constraint is a logical relation among several variables which has a value in a certain domain. Constraints impose some restrictions to Optimizer for proper resources reallocation.
Search heuristics is a set of heuristics that enables the Optimizer to find the satisfactory solutions quickly. Using Search heuristics we can find a solution in a polynomial time O(n²), where n being the number of servers.
Finally, the output of Optimizer is VM reallocation or their distribution over the servers. This output is then translated into actions such as moving VMs and powering on/off servers.
The FIT4Green Optimizer engine is based on the open-source library Entropy (a joint effort between the Ecole des Mines de Nantes and the University of Sophia Antipolis, http://entropy.gforge.inria.fr). Entropy is an autonomous virtual machines manager for hosting platforms. It provides an API to reconfigures the state and the placement of the VMs according to high-level constraints. The flexibility of the reconfiguration algorithm is provided by the usage of Choco (http://choco.emn.fr), a CP solver. In FIT4Green, we exploited the extension mechanisms of Entropy by providing our own constraints, heuristics and power objective in order to aim for energy consumption saving.
FIT4Green – Experimentation and validation
FIT4Green aims at developing and deploying a set of global energy optimization policies for data centres, applicable to and verifiable for different computing styles (traditional, super-computing and cloud).
The project assesses the effectiveness of those policies in saving energy by testing and evaluating the components implementing the global energy optimization policies in three different test beds in real trial environments, each one representative of a specific computing style, for single or federated data centres, respectively. Additionally FIT4Green considers the lack of power aware management features from most networking equipment pieces by providing a general energy consumption model of network outfit, enabling to take into account the network impact on global energy consumption of a data centre federation.
The approach of these tests will follow the spiral lifecycle models: three phases of pilot tests will iteratively refine and add complexity to the models and tools specifically developed to optimize the energy consumption and GHG emissions of the different scenarios. The three different phases of the test pilots:
- In a first release, the tests assess results related to single data centre optimizations by monitoring the effectiveness of the ‘FIT4Green Plug-In inside Single Data Centres’.
- In the second release, the tests revise single data centre optimizations and the first release of federated data centre by monitoring the effectiveness of the ‘FIT4Green Enhanced Energy Control Plug-In inside Single Data Centres and Control Desk for Federated Data centres’.
- In the last, the tests release final results from both single and federated data centre optimizations by monitoring the effectiveness of the ‘FIT4Green Full-Featured Enhanced Control Plug-In and Control Desk for Federated Data Centres’.
All three test phases are carried on in the three test beds representing different computing styles: ENI for traditional enterprise IT environments, JÜLICH for High Performance Computing, HPIS for cloud IaaS. A fourth testbed on network infrastructure is instrumented by ICL in the first phase only, to provide inputs for the second and third cycles.
For specific information about the testbeds environment and configuration, testing methodology, numerical results, and technical and usability evaluation related to the first tests execution release please refer to [link to D6.2]
FIT4Green has developed a plug-in prototype so as to demonstrate the feasibility of the research thesis within the project: “for a data centre with no previous steps with regard to energy optimization, FIT4Green policies and models can provide on average 20% savings in direct server and network devices energy consumption and induce an additional 30% savings due to reduced cooling needs.”
The goal of the developed plug-in is to dynamically optimize deployment of the applications and services hosted and running at a single- or federated-site data centres in order to minimize the energy consumption or carbon (CO2) emissions, that is, trying to consolidate load in such a way that some hardware resources can be turned off.
The plug-in is based on energy consumption models and energy optimizing policies ; the former are based on energy related information regarding the infrastructure and hardware, including hardware specific energy optimization features while the latter capture knowledge and strategies that provide an active means of improving energy efficiency.
The plug-in is capable of interfacing and interacting with management and automation tools operational at data centres. Thereof, it provides access to the existing monitoring systems of the data centre so as to obtain load related data that can then be used to generate realistic energy consumption models of the data centre. Furthermore, the plug-in allows selecting and enforcing energy optimizing policies for application and service redeployment.
The energy control plug-in is being tested and evaluated in specific trial sites for each one of the different computing styles.
