Cost-Effective Feature Placement of Customizable Multi-Tenant Applications in the Cloud

Abstract

Cloud computing technologies can be used to more flexibly provision application resources. By exploiting multi-tenancy, instances can be shared between users, lowering the cost of providing applications. A weakness of current cloud offerings however, is the difficulty of creating customizable applications that retain these advantages. In this article, we define a feature-based cloud resource management model, making use of Software Product Line Engineering (SPLE) techniques, where applications are composed of feature instances using a service-oriented architecture. We focus on how resources can be allocated in a cost-effective way within this model, a problem which we refer to as the feature placement problem. A formal description of this problem, that can be used to allocate resources in a cost-effective way, is provided. We take both the cost of failure to place features, and the cost of using servers into account, making it possible to take energy costs or the cost of public cloud infrastructure into consideration during the placement calculation. Four algorithms that can be used to solve the feature placement problem are defined. We evaluate the algorithm solutions, comparing them with the optimal solution determined using an integer linear problem solver, and evaluating the execution times of the algorithms, making use of both generated inputs and a use case based on three applications. We show that, using our approach a higher degree of multi-tenancy can be achieved, and that for the considered scenarios, taking the relationships between features into account and using application-oriented placement performs 25% to 40% better than a purely feature-oriented placement.

Allocating resources for customizable multi-tenant applications in clouds using dynamic feature placement

“…This article extends our previous work related to feature placement [4,5], which focused on the static feature placement problem, and describes and evaluations new dynamic feature placement algorithms that can be used in a context where applications are added and removed through time. The modeling approach we use is further based on our work on feature model conversion [3], which focuses more on how the code modules themselves are defined and how customizable applications within our approach can be designed rather than on how these modules are managed at runtime, which is the focus of this article.…”

“…Figure 3 shows an illustrative example a simple feature model consisting of six features. The relations between the various features are expressed using the Pure::Variants notation [31] which we also used in our previous work [3,5]. This model corresponds to three relations: Optional(a, b), Or(a, {c, d}), and Alternative(c, {e, f}).…”

“…We previously presented a formal problem formulation for the static feature placement problem in [5]. In this section we briefly describe the problem model, focusing on the variables that are needed to add migration-awareness the model.…”

“…In this section we briefly describe the problem model, focusing on the variables that are needed to add migration-awareness the model. For a more detailed discussion of the model parameters and formulation we refer to [5]. We will end this section by discussing how migration-awareness can be added to the model to make it useful for dynamic application placement scenarios.…”

“…be cheaper to use an existing high-performance instance for a tenant that does not pay for such an instance rather than to allocate a low-performance instance specifically for this tenant. We previously addressed resource allocation taking this information into account, referred to as feature placement, in [4] and [5], but the approach however resulted in a static resource allocation, that has to be recomputed periodically. In doing so, the number of migrations is not taken into account, which adversely impacts the performance of the system when services are migrated.…”