Search

Function-as-a-Service (FaaS) has become a mainstream cloud computing paradigm for developers to build cloud-native applications in recent years. By taking advantage of serverless architecture, FaaS applications bring many desirable benefits, including built-in scalability, high availability, and improved cost-effectiveness. However, predictability and trade-off of performance and cost are still key pitfalls for FaaS applications due to poor infrastructure transparency and lack of performance and cost models that fit the new paradigm. In this study, we therefore fill this gap by proposing formal performance and cost modeling and optimization algorithms, which enable accurate prediction and fine-grained control over the performance and cost of FaaS applications. The proposed model and algorithms provide better predictability and trade-off of performance and cost for FaaS applications, which help developers to make informed decisions on cost reduction, performance improvement, and configuration optimization. We validate the proposed model and algorithms via extensive experiments on AWS. We show that the modeling algorithms can accurately estimate critical metrics, including response time, cost, exit status, and their distributions, regardless of the complexity and scale of the application workflow. Also, the depth-first bottleneck alleviation algorithm for trade-off analysis can effectively solve two optimization problems with fine-grained constraints. [ABSTRACT FROM AUTHOR]

In this paper, we evaluate the performance of cloud centers with high degree of virtualization and Poisson batch task arrivals. To this end, we develop an analytical model and validate it with an independent simulation model. Task service times are modeled with a general probability distribution, but the model also accounts for the deterioration of performance due to the workload at each node. The model allows for calculation of important performance indicators such as mean response time, waiting time in the queue, queue length, blocking probability, probability of immediate service, and probability distribution of the number of tasks in the system. Furthermore, we show that the performance of a cloud center may be improved if incoming requests are partitioned on the basis of the coefficient of variation of service time and batch size. [ABSTRACT FROM PUBLISHER]

Accurate performance evaluation of cloud computing resources is a necessary prerequisite for ensuring that quality of service parameters remain within agreed limits. In this paper, we employ both the analytical and simulation modeling to addresses the complexity of cloud computing systems. Analytical model is comprised of distinct functional submodels, the results of which are combined in an iterative manner to obtain the solution with required accuracy. Our models incorporate the important features of cloud centers such as batch arrival of user requests, resource virtualization, and realistic servicing steps, to obtain important performance metrics such as task blocking probability and total waiting time incurred on user requests. Also, our results reveal important insights for capacity planning to control delay of servicing users requests. [ABSTRACT FROM AUTHOR]

In this paper, we propose an analytical performance model that addresses the complexity of cloud centers through distinct stochastic submodels, the results of which are integrated to obtain the overall solution. Our model incorporates the important aspects of cloud centers such as pool management, compound requests (i.e., a set of requests submitted by one user simultaneously), resource virtualization and realistic servicing steps. In this manner, we obtain not only a detailed assessment of cloud center performance, but also clear insights into equilibrium arrangement and capacity planning that allows servicing delays, task rejection probability, and power consumption to be kept under control. [ABSTRACT FROM AUTHOR]