Enriching Cloud-native Applications with Sustainability Features

From a research perspective, two complementary approaches have been proposed for reducing the environmental impact of data centers: (i) design efficient facilities [2] and (ii) improve server utilization [3].

It will become essential that applications are also designed and executed based on their environmental footprint. At the same time, an evolution in the architectural style of the applications has been observed, moving from monolithic to cloud-native applications, designed to take advantage of the characteristics of cloud computing [7] [8] and becoming the de facto standard in industry [9].

We aim to increase the sustainability awareness of the stakeholders involved in cloud-native application management, from design to deployment, and to support the development of sustainable applications by enriching them with metadata to be exploited for more efficient deployment. Starting from the general Sustainable Application Design Process (SADP) [10], we present an architecture supporting the methodology and define the components needed for its enactment. Additionally, we implemented and tested a prototype on two use cases.

A more comprehensive approach to sustainability in cloud computing has been presented by Gill and Buyya [15], through a set of taxonomies for several aspects of Cloud Data Centers sustainability taking the cloud providers’ perspective. The considered aspects include application design, sustainability metrics, capacity planning, energy management, and virtualization.

The cloud provider perspective is also adopted by exploiting the intermittent availability of renewable energy by Thi, Pierson, and Da Costa, Hu, Li, and Sun [16], [17], involving operations such as server consolidation [3], [18] and shutdown policies to ensure a trade-off between energy efficiency and performance [19].

Energy reduction should be a shared responsibility between the cloud provider and the application provider (or cloud consumer) [22]. However, application providers usually lack information to improve their environmental impact [23] while, on the other hand, cloud providers see the applications they host as black boxes [11].

Limited attempts have been made to include sustainability in the application design of Green Information Systems (Green IS) by taking the application provider’s perspective. An empirical approach towards energy efficiency of applications in the cloud from the application provider perspective, focusing on architectural aspects, is discussed by Vos, Lago, et al. [23]. Vos, Lago, et al. classify tactics into three categories: resource monitoring (workload energy assessment), resource allocation (scaling, scheduling, brokering), and resource adaptation (data reduction, granular scaling, batch execution, edge deployment).

Current research mainly focuses on dynamic resource allocation and scheduling according to energy efficiency optimization [24], [25].

Other attempts have proposed general principles and lack practical solutions [28]–[31].

Microsoft Azure [32], Google Cloud Platform [33], and Amazon Web Service [34] are now providing a dashboard to report the energy consumption and CO2 emissions at the cloud service level. All these tools lack details and do not provide timely information.

Third parties tools have been developed to timely estimate applications’ energy usage and carbon emissions. The CodeCarbon initiative [35] estimates CO2 emissions of Machine Learning tasks in the geographical location and the energy mix of the country in which the application is deployed. Similarly, The Cloud Carbon Footprint Tool [36] estimates the energy consumption of cloud instances. Power is also one of the metrics considered by Brondolin and Santambrogio, providing closed-box monitoring for multi-component applications [37]. Some proposals aimed to estimate the energy efficiency of specific applications in embedded systems [38] but cannot be applied in complex cloud infrastructures. A virtual energy system exploited to optimize the carbon efficiency of applications is proposed by Souza, Bashir, et al. [39].

Cloud-native applications empower the design and execution of scalable, loosely coupled, resilient, manageable, and observable applications [8]. These applications generate complex workflows as a consequence of the microservice architectural style. Specific characteristics of these applications can be exploited to improve energy efficiency.

Elasticity can be used to balance energy and performance [41]. Energydriven considerations in Mobile Cloud Computing can lead to microservices’ offloading [42]. An attempt to consider the workflow and the interdependencies between microservices of the same applications is presented by Saboor, Mahmood, et al. [43], where a ranking approach allocates microservices to containers according to their interactions.

The design of adaptive applications, able to change their workflow according to the current workload and energy mix, is necessary to empower green applications from an application provider perspective.

The power-aware brownout strategy proposed Papadopoulos, Krzywda, et al. presented an approach to adapt the workflow execution by producing a response containing minimal components (mandatory part) while disabling additional components in case of limitations in power availability [44].

Gerostathopoulos, Raibulet, and Lago suggest the usage of decision maps to make sustainability-driven decisions [45]. Decision maps associate tactics and sustainability goals. For example, they use service scaling and something similar to brown out to adapt the application workflow from an economical and technical sustainability perspective.

We opted to represent microservice choreography through Business Process Model and Notation (BPMN) fragments, exploiting a well-known process modeling notation to represent microservices interactions [46].

The proposed approach considers three actors: (i) the application designer is made aware of the current sustainability of the application through a set of metrics and indicators and is in charge of setting sustainability targets and of re-designing the application adding sustain- ability features; (ii) the application developer is in charge of implementing the sustainability features included by the application designer according to the specific framework used for the application development and management; (iii) the infrastructure provider adapts the application deployment and scheduling according to the rules and characteristics added in the design and development phases and provides energy and carbon footprint data about the current application execution.

The SADP approach described in [10] identifies a set of features and makes them explicitly defined during the design process. However, these features need to be visible to the infrastructure provider. Thus, the features added at design time must be reflected in the implementation of the application components. This paper introduces a Sustainable Application Design Architecture (SADA) and focuses on the synergy between the design and the development phases.

Sustainable Application Design Process (SADP):

• Sustainable Awareness: requirements and capabilities → assign correct amount of computational resources and best deployment location

• Microservice Classification: optional or mandatory → application workflow can be adapted according to the current execution context or if there is a green energy sources shortage

• Microservice Enrichment: multiple implementations with different functional and non-functional properties → adaptive workflow

The architecture is characterized by three main subsystems, each one responsible for one of the phases of the application management and addressed to one of the actors defined in Section III:

• Sustainable Application Design, addressed to the application designer;
• Web Service Development Framework, addressed to the application developer;
• Sustainability-Aware Deployment System, addressed to the infrastructure provider.

The Sustainable Application Design subsystem has to: (i) enable the (re)design of the application, including the features of the SADP; (ii) provide feedback on the current sustainability level of the application; (iii) enable to export the application model in a machine-readable format to ease the implementation of the defined characteristics.

We propose to embed the sustainability context embedded in the BPMN notation [48], a standard to model an organization’s internal processes and collaborations and orchestrations between different organizations. Following the approach adopted in [46], the application can be represented as a set of microservices orchestrated by a general process, defining their order of execution (workflow).

The Business Process Modeler component implements the front-end for the application design using BPMN. It enables the application designer to load and edit an existing model or to create a new one. It also allows the enrichment of this model with the sustainability features introduced in Section III. Each task in the BPMN model represents a microservice of the application and is linked to its implementation by specifying the service identifier. This enables the alignment between the application representation and its actual implementation.

The Web Service Development Framework subsystem defines how to enrich the microservices composing the application with the metadata expressed in the design. It also enables the application designer and developer to exchange sustainability information.

The Carbon-Aware Deployment System is the connection point between the Infrastructure Provider, who is in charge of deploying the application, and the Application Designer and Developer, who designs the adaptive workflows.

We instantiated the web service development framework integration components into a DSL-based Apodini web service application framework. Apodini is a declarative Swift framework to express the interface, requirements, and functionalities of web services in a single internal DSL independent of any specific middleware, protocol, or Interface Description Language (IDL) [50], [51].

We demonstrate the implementation of the architectural components using a Flight Booking application shown in Figure 3). The application consists of five microservices. The Weather Information and the Rental Car Booking components are optional (their execution can be skipped if needed), while

Microservices of the Flight Booking application

for the Flight Search component, three versions have been defined: i) a normal version, collecting updated data from flight companies at each new request; ii) a low-power version using cached and possibly outdated information in the preliminary search phases to reduce energy waste; iii) a high-performance version, using a recommender system to improve the results presented to the customer and their order.

Example Online Boutique:

• two tasks have been labeled as optional: Advertisement and Product Recommendation
• two tasks have been complemented with alternative versions: a low-power version Product Catalog has been configured to show a limited set of products during the navigation, the Product Recommendation component recommends a reduced set of products using a simpler algorithm

At the current stage, it is impossible to quantify the benefits of the proposed approach from an environmental sustainability perspective. The actual reduction in energy and emissions depend on several factors, including the involvement of all the stakeholders. However, some preliminary tests have been executed to compare the emissions generated by the applications presented as examples under different execution modalities. In a best-effort approach given the described constraints, we compared the online Boutique application introduced in Section V-C execution of the out-of-the-box application with the execution of the same application with the optional and low-power version of its tasks simulating a workload of 100 users. The results show a reduction of 27% in computational resource demand.

However, the presented sustainable design can only affect the environmental impact of cloud-native applications if the synergy between all the stakeholders is enacted and appropriate actions are taken accordingly in their management at runtime.