Time-preference-based on-spot bundled cloud-service provisioning

Highlights

• A decision support framework to incorporate customer time preference in cloud service provisioning.

• Offer cloud services as a bundle.

• Model customer behaviour to maximize cloud provider revenue.

• Revenue Time Trade-Off.

• Potential implication in cloud brokering.

Abstract

The cloud computing spot instance is one offering that vendors are leveraging to provide differentiated service to an expanding pay-per-use computing market. Spot instances have cost advantages, albeit at a trade-off of interruptions that can occur when the user's bid price falls below the spot price. The interruptions are often exacerbated since customers often require resources in bundles. For these reasons, customers might have to wait for a long time before their jobs are completed. In this paper, we propose a behavioral-economic model in the form of time-preference-based bids, wherein users are willing to use and bid for services at other times if the vendor cannot provide the resources at the preferred time. Given such bids, we consider the problem of provisioning for such service requests. We develop a time-preference-based optimization model. Since the optimization model is NP-Hard, we develop rule-based genetic algorithms. We have obtained very encouraging results with respect to standard commercial solver as a benchmark. In turn, our results provide evidence for the viability of our approach for online service-provisioning problems.

Introduction

Internet-based real-time services have been eliciting interest among customers in recent years, owing, in part, to the timely, convenient, and cost-effective delivery of services. One such service is cloud computing¹ [1]. The revenue from the worldwide public cloud market is expected to be $306.9 and $364 Billion by the end of 2021 and 2022 respectively [2]. Cloud services are gaining popularity, as many enterprises are moving their operations to the cloud. For example, Expedia Inc. has migrated 80% of its services to Amazon Cloud to execute its computation-intensive tasks [3], where Expedia requires different types of resources. Thus, to provide differentiated services, cloud companies are also offering different types of services. For example, Amazon Web Service [4] offers several types of cloud-enabled services: On-Demand, Reserve, Scheduled, Dedicated and Dedicated Host, and different spot-market variants. We focus on spot-market instances in this paper.

Spot market is an offering where the service provider's unused capacity is made available to the customers without any commitment, often at significantly lower price (sometimes 90% cheaper) as compared to On-Demand and other markets [5]. Due to such discounts, spot instances are used by organizations flexible about run-time and interruptions. For example, digital platforms such as AdRoll uses Amazon EC2 spot instances on a long-term basis for their ad-delivery to 10,000 advertisers across 100 countries; see [6] for such examples.

Spot-market works on users' bids, and if that value is greater than the current spot-price,² the service is assigned to the customer. Customers have no control over the spot-price. Hence, in contrast to other service types, the provider can terminate the instance whenever the user's bid falls below the current spot-price [7]. Such ad hoc service terms may pose an issue for time-sensitive customers, who might get dissatisfied with such unforeseen interruptions and potentially switch to competitors. This makes cloud migration challenging. Based on an analysis of cloud industry, an earlier work by Marston et al. [8] (p. 186, 2011) suggested that: “...the broad IS research agenda in cloud computing can be divided into six main potential streams [including] cloud computing economics”. As such, in this paper, we take a business/economic perspective and consider how customer preferences must be factored-in to provision them with spot-instances. We develop a service-provisioning algorithm to maximize the provider's revenue, under three key aspects:

The first aspect of our research is bundling. A bundle specifies the relative time intervals at which a group of related cloud services are required as a package for the customer-job's execution. Often, cloud users require multiple heterogeneous resources to execute their tasks (see, for example, [[9], [10], [11], [12], [13], [14]] for works that consider bundling). As indicated by Shi et al. [15] (p. 72, 2014), task decomposition of user-jobs implies that: “A cloud computing job in practice often demands a bundle of heterogenous VM instances for its successful execution”. For example, in solving complex molecular dynamics algorithms connected with biological experiments, a user requires multiple instances of different types of cloud virtual machines (based on virtual CPU, memory, storage capacity etc.) in a particular sequence. To demand all these virtual machines for the entire time-horizon would be sub-optimal for the user. Instead, therefore, the user distributes the resource requirements across the time horizon and requests the cloud provider to allocate the resources as a bundle, so that they can plan ahead. Fig. 1 depicts a similar scenario where User 1's bundle requirement includes M1-M2-M3 at time t, and M1-M2 and M1, respectively, at one and three time-periods after t (see Fig. 1³) as a bundle. Likewise, User i also submits such a bundle-request.

The second aspect of our research seeks to mitigate the time-sensitivity issue mentioned above. For this, we consider the customer's preference for task-completion time. Previous cloud-related literature has also considered time-sensitivity (e.g., [[16], [17], [18], [19]]), by looking at it from a scheduling perspective, employing scheduling rules based on job-deadlines (e.g., earliest deadline first). In contrast, we consider time-sensitivity by expanding the time-horizon of the bids (from a single period to multiple periods). Within such an expanded time-horizon, we consider the customer's time-preference (TP)⁴ by using a behavioral economic model; i.e., we seek the completion-time preference vis-à-vis the bid-price the customer is willing to pay (i.e., their valuation) for that preference (in general, faster deliveries mean higher bids). Thus, in cases where the provider cannot offer the service to the user immediately due to lack of cloud-resource availability, the user-specified TPs can be used to satisfy the demands starting at a succeeding time-period. For example, as shown in Fig. 1, User 1 is willing to pay 25 if the bundle is accommodated starting time-period t, but only 24 for an allocation at period t + 1; the corresponding numbers for User i are 24 and 22.

The third aspect of our research is the need for a real-time solution for spot-market cloud service provisioning. Unlike the other types of cloud-services (e.g., reservation type), wherein there is sufficient time to make decisions, the decision-making time for spot-markets is low. Despite this requirement, due to limited service availability, the provider takes a longer time to offer a requested service to the customer. In many cases, the customer does not get the resource even after waiting for a long time. From the customer's perspective, a lengthy decision time could force the customer to leave for other competitors. From the provider's perspective, the supply, demand, and prices are dynamic, and hence the allocation is relevant only if it is done quickly. Thus, the service-provisioning decision based on customer's TP must be done in a short time (less than 10 min following AWS standard—see https://aws.amazon.com/ec2/faqs/). Given the three aspects of the problem mentioned above, we aim to develop a service provisioning algorithm to maximize the cloud provider's revenue.

As we will see in Section 2, the consideration of these three aspects is not there in the literature. We look into the TP literature to incorporate customer preference, along with resource bundling. This bridges the current research gaps in the cloud-service-provisioning domain. We first develop a binary integer-programming optimization model to maximize the overall revenue of the cloud provider. However, due to the hardness of the problem (see Theorem 1), we come up with a genetic-algorithm-based meta-heuristic approach to solve the problem within the stipulated time. Finally, to test the practicality of our approach, we demonstrate some key features of the algorithm with simulated datasets based on real-life scenarios.

In this work, we make three significant contributions: (1) We develop a new way of allocating cloud resources considering customer's TPs; this approach has the potential to generate better revenue than the existing approaches (2) We incorporate service bundles along with TP into our approach, bridging the gap between behavioral economics, and the practice and research of cloud service management. (3) Our meta-heuristic approach contributes to providing a better feasible solution than what general-purpose commercial software provides within the stipulated decision-time.

We first present the related work in Section 2, followed by the model development and solution approach in 3 Model development, 4 Solution approach respectively. Next, we discuss the experimental design and results in Section 5. Section 6 provides the research implications and insights of our work. Finally, we conclude our paper in Section 7.