Event-driven approach for predictive and proactive management of SLA violations in the Cloud of Things

Highlights

• Utilizes an event-driven approach to model the future state of SLOs and ascertain its impact on the compliance of the SLAs to proactively manage their violations.

• Utilizes event calculus to model each event into its fluents to determine the future state of a SLO and the compliance of the SLA to defined agreement.

• Utilizes Bayesian Networks to model the future state of uncertain events.

• Validates the proposed methodology on a real-world cloud dataset to demonstrate the applicability of our proposed model.

Abstract

In a dynamic environment such as the cloud-of-things, one of the most critical factors for successful service delivery is the QoS under defined constraints. Even though guarantees in the form of service level agreements (SLAs) are provided to users, many services exhibit dynamic Quality of Service (QoS) variations. This QoS variation as well as changes in the behavior and state of the service is caused by some internal events (such as varying loads) and external events (such as location and weather), which results in frequent SLA violations. Most of the existing violation prediction approaches use historic data to predict future QoS values. They do not consider dynamic changes and the events that cause these changes in QoS attributes. In this paper, we propose an event-driven-based proactive approach for predicting SLA violations by combining logic-based reasoning and probabilistic inferencing. The results show that our proposed approach is efficient and proactively identifies SLA violations under uncertain QoS observations.

Introduction

Cloud computing and the Internet of Things (IoT) are two emerging paradigms that have enabled businesses to access on-demand computing resources and be connected to various computing devices spread across different locations, respectively. The advantages of merging cloud computing and IoT have given rise to the notion of the cloud-of-things (CoT) [1]. One of the main challenges for successful service delivery in the CoT environment is the quality of service (QoS) under defined constraints, such as response time range, minimum service availability or a certain throughput. These service constraints are usually stated by the provider in the service offer and are mutually agreed upon by the provider and consumer in the service level agreement (SLA). However, due to the characteristics of the CoT and the services themselves, many services exhibit dynamic QoS variations that result in frequent changes in their behaviors leading to SLA violations [[2], [3]]. For example, some QoS attributes such as response time, service availability, and operational cost can change dynamically due to various internal (such as varying loads) and external (such as location and weather) events [4]. This leads to problems at two stages as follows:

Stage 1: Most of the QoS representation approaches that represent a service’s QoS specifics are limited to showing their static QoS attributes only [[5], [6]]. In other words, they are unable to dynamically represent a change in the service quality parameters to users on which the SLAs are formed. Hence, service users do not have the correct platform to make service selection decisions. In order to further understand the importance of the issue of dynamic QoS variations, consider that there is more than one service with a similar functionality. Then the critical factor for choosing the best service comes down to the QoS attributes, which is not very well defined. So the services that are chosen based on using only these static QoS constraints might not be the best ones [7].

Stage 2: To model the dynamic change in a service’s QoS parameters after it is provisioned, apart from monitoring, existing approaches use the service’s historic QoS values and predict future ones [[8], [9]]. However, in a dynamic service amalgamated environment such as the CoT, a resulting service may be a combination of sub-services. The existing approaches do not proactively consider the dynamic changes and/or the events that cause such changes in the resulting service’s QoS attributes to manage service violations.

To emphasize the importance of each stage and their drawbacks, Fig. 1 shows the service management life cycle that has pre-SLA formation (referred to as pre-interaction in the remainder of the paper) and post-SLA formation (post-interaction) time phases. The pre-interaction time phase includes the tasks of service representation, discovery, negotiation and selection stages while the post-interaction time phase includes service provisioning, monitoring and violation detection. Both time phases have general monitoring and adaptation phases to ensure better service management and QoS delivery to the customers. In the context of service management and as shown in Table 1, it is important to address the drawbacks of stage 1 for the accurate representation of a service in the phases of service representation, service discovery, service selection and negotiation; whereas it is important to address the drawbacks of stage 2 for the better management of a service during the phases of service provisioning and violation detection.

Our focus in this paper is more towards stage 2, as the limitations in the current literature demand the development of an event-driven approach that can identify events which will impact a service’s QoS and model their effect on QoS attributes to manage SLA violations. Event-driven computing [10] is about defining an automatic reaction to events. An event can be defined as a happening within a system or domain. A raw (simple) event is produced directly from the provider whereas derived events are produced by some processing agent that applies logic on a set of input (simple) events and produces a set of output derived (complex) events [11]. Event-driven computing can be used in two ways to respond to events: first, responding to events that have occurred, i.e. a reactive response and second, responding to events that may occur, i.e. a proactive response.

In this paper, we propose a proactive event-driven approach for predicting and managing SLA violations. Proactive event-driven computing for managing SLA violations evolves from reactive event-driven computing and involves dealing with uncertainty at two levels. The first level of uncertainty is determining events that will lead to a variation in the defined QoS and the second type of uncertainty is determining or predicting what will happen as a result of these events occurring. In other words, proactive event-driven computing predicts the occurrence of events and the uncertainty of the situation (or effect) after the occurrence of an event in order to reduce the negative consequences and exploit future prospects that would otherwise be missed by avoiding SLA violations. Based on this, our proposed approach first triggers the defined rules to ascertain the event that causes QoS variation and then predicts a possible situation, such as an SLA violation due to the occurrence of the event using logical reasoning and probabilistic inferencing. The rest of the paper is organized as follows. In the next section, we discuss the related work in the area of SLA violation. The proposed framework is presented in Section 3. In Section 4, we formalize the WS-Agreement [12] with respect to the proposed framework. The syntax and semantics of the proposed system are presented in Section 5. In Section 6, we describe the reasoning and decision making part of our proposed framework, followed by the experimental validation of the proposed system in Section 7. Finally, Section 8 concludes the paper.