Article Preview
Top1. Introduction
Cloud Computing has turned into a consolidated approach for delivery of many end-to-end services through on-demand provisioning of virtualized resources in the pay-as-you go system model. In this scenario, a growing number of companies are adopting public cloud computing services: Infrastructure-as-a-Service (IaaS), which offers a relatively low level virtualized hardware, Platform-as-a-Service (PaaS), which is a development platform layered on top of IaaS resources and provides higher-level platform-specific APIs, and Software-as-a-Service (SaaS), which is an implementation of specific applications with cloud capabilities.
Cloud customers outsource their computational demand to public providers and pay accordingly. Specially, regarding the infrastructure as a service, costumers lease storage space, processing power, and communication from IaaS providers (Amazon EC2, Google Compute Engine, ElasticHosts, CloudSigma, and JoyentCloud). The most common strategy to delivery infrastructure services is in the form of service level agreements (SLAs) of virtual machines (VMs), where users can deploy distributed applications.
Customers choose the VMs profiles (SLAs) offered by the provider, and pay according to how long these VMs are utilized (hourly, monthly, semi-annual, and annual, for instance). IaaS resources specified by SLAs offer a variety of properties, from which the user can choose the best fit for his/her needs. Moreover, depending on the application demands, users can rent resources from different providers. By deploying cloud services across several cloud providers instead of using a single one, users can have benefits like cost reduction, load balancing, and better fault tolerance, and also avoid vendor lock-in. Public IaaS providers offer different pricing schemes and charging models. For example, Amazon EC2 offers three types of charging models: On-Demand (OD), Reserved (RE), and Spot (SP). On-demand instances are available by request at a given price, so as the reserved instances. However, to have the right to use reserved instances, which present lower prices than on-demand instances, the user has to pay a fee for a term (a year, for instance). While on-demand and reserved instances are only interrupted at the users will, the IaaS provider can interrupt the cheaper spot instances as they are rented via an auction system at lower prices.
In general, consumers seek to pay less for the better available service (Houidi, 2011; Malawski, 2013; Vieira, 2013; Li, 2013; Assunção, 2010; Vieira, 2015). On the other hand, cloud providers have the objective of attending the most clients with the least resources in order to increase profit. Moreover, if besides the public clouds a private cloud is also available, the consumer can have difficulties in determining the best combination to fulfill application requirements and reduce costs.
Motivated by the challenges encountered by the users when choosing IaaS providers and VM charging models to schedule their applications and, at the same time reducing monetary costs, we introduced in (Vieira, 2014) a mechanism based on redundancy from a mixed utilization of reserved and spot virtual machine (VM) instances. This mechanism allows balancing between cost and availability, while reducing the chances of a QoS violation. This is achieved by utilizing cheaper machines with lower availability guarantees, such as spot machines at Amazon EC2, to allocate requests that accept a certain level of risk to be preempted, while ensuring QoS through redundant concurrent execution in more expensive instances.
The strategy to decrease the cost of VM requests scheduling presented by Vieira et al. (Vieira, 2014) considers solely the charging models in the SLA. However, many other characteristics in the SLA must be taken into account when choosing a VM to execute users applications: processing capacity, bandwidth, storage, isolation, etc. Such characteristics define a VM type, which is offered through an SLA to the IaaS client.
In this paper, we extend the proposal from (Vieira, 2014) to introduce a new dimension to the SLA in order to deal with characteristics that go beyond the charging model. In this sense, a two-dimensional SLA better fits the user needs and allows proper utilization of resources available in IaaS providers.
We consider that VM requests arrive to a broker belonging to a hybrid cloud or to an SaaS or PaaS provider. This broker allocates requests into VMs in public clouds. In this context, the main contributions of this work are: (i) An SLA with two dimensions, namely charging model and VM type; (ii) an integer linear program (ILP) to compute the scheduling regarding an SLA with two dimensions; and (iii) analysis of experimental results.