Task Feature Based Trading Strategy For Resource Efficiency In Public IaaS Clouds

Public Infrastructure-as-a-Service cloud computing services provide computing resources to tenants in the form of virtual machine(VM)instances and charge them with a pay-as-you-go fashion.By executing computing tasks in the cloud,tenants not only avoid the huge in-vestment in the IT infrastructures but also can easily scale their resource usage up and down according to the workload.On the other side,the service provider can make a profit with a lowered service cost,which is enforced by multiplexing resource usage among tenants and the economics of scale obtained through hosting a huge number of cloud applications in a virtualized datacenter.With the current trading strategies,both the provider and tenants are faced with the challenge of further enhancing the profit gained in public cloud services.From the perspec-tive of providers,the cloud datacenter is built with an enormous resource capacity to handle the demand peaks of tenants,while a large portion of it would be idle in demand valleys,impairing the resource efficiency.Moreover,tenants tasks typically cannot fully utilize the resources in the VM instances,resulting in resource fragments that cannot be efficiently recycled without violating service level agreements,thus making the resource inefficiency even severer.For the tenants,the resource fragment problem would waste their payment.In addition,existing trading mechanisms for transient idle resource requires complicated deci-sion making on instance renting and do not have any guarantee on performance,therefore,making it very difficult for tenants to exploit cheap idle resource and save cost.The above mentioned issues would hinder the adaption of public cloud services for a wide range of applications,thus becoming one of the main obstacles in its developing progress.The pursuit of cost saving and profit gaining for the provider and tenants is actually cor-related.With carefully designed trading mechanisms,the provider could incentivize tenants,with price discounts,to help improve resource efficiency according to their tasks features.In the meanwhile,the tenants can save their cost with the given discount.By combining re-source allocation methods in economics with resource management techniques in the cloud,three public cloud services,i.e.,flexible instance,elastic instance,and group buying in-stance is proposed,with joint consideration on the demands from both the service provider and tenants.To mitigate the pressure of demand peaks on the cloud datacenter,flexible instance exploits the scheduling flexibility of delay tolerant tasks and schedules them into non-peak periods with price discounts.To ensure that tenant jobs can finish within their respective deadlines,the Flexible instance introduce the concept of execution ratio,which is a metric of job scheduling flexibility,as a novel pricing factor and scheduling constraint.Flexible instance uses a two-stage framework to maximize provider revenue.The first strategy re-serves sufficient resources for tasks to meet deadlines and uses a dynamic pricing mechanism to calculate the payments according to a price curve,so as to cope with the demand-supply relationship of resources.The second stage leverages the Nash bargaining framework to con-duct task scheduling and pricing in non-peak periods.The resulted scheduling and pricing scheme can achieve resource efficiency,proportional fairness and provider revenue maxi-mization simultaneously.Based on the elasticity of resource demand in tenant tasks,elastic instance improves resource efficiency through incentivizing tenants to consume more resource in demand val-leys.To simplify decision making for tenant on resource renting and avoid the impact of their local optimization strategies on the global resource efficiency,elastic instance only requires tenants to describe their demand elasticity and let the provider conduct global re-source allocation.Elastic nstance formulates the problem of social welfare maximization in resource allocation and exploits the ADMM(Alternating Direction Method of Multiplier)al-gorithm to perform efficient solution calculation.In the pricing mechanism,elastic instance uses the celebrated VCG(Vickrey-Clarke-Groves)strategy in auction theory to ensure the truthfulness of tenant submitted resource demand.To avoid fragmented resource,group buying instance leverages the idea of group buy-ing and designs a trading strategy for resource efficient task placement.In group buying instance,tenant tasks with complementary resource demands are organized into groups to collectively buy resources in predefined group buying deals.To enhance resource utilization of both newly created group buying deals and existing deals,group buying instance utilizes static and dynamic strategies for task placement.The static strategy packs a batch of tasks into newly created group buying deals,and adapts the branch-and-price framework to design the optimal algorithm as well as a polynomial time approximation solution.The dynamic strategy allocates newly arrived tasks into the idle resource of existing deals,and proposes a learn-and-pack algorithm along with the theoretical analysis on its competitive ratio.The design and performance analysis on the above mentioned trading mechanisms demonstrate the importance of trading mechanism as the connection between the provider and tenants in public cloud services.A smartly designed trading mechanism tenants could help the provider with resource efficiency enhancement,meanwhile enjoying a price dis-count for task execution.