Modeling Service Applications for Optimal Parallel Embedding

作     者：Huang, Changcheng Zhu, Jiafeng 

作者机构：Carleton Univ Dept Syst & Comp Engn Ottawa ON K1S 5B6 Canada Huawei Technol Inc Santa Clara CA USA 

出 版 物：《IEEE TRANSACTIONS ON CLOUD COMPUTING》 (IEEE Trans. Cloud Comput.)

年 卷 期：2018年第6卷第4期

页      面：1067-1079页

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0835[工学-软件工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

主　　题：Architecture distributed network network topology distributed application modeling technique distributed programming 

摘      要：Leveraging the traditional virtual network concept, some recent research works have proposed the Virtual Data Center (VDC) as an abstraction to capture both bandwidth and compute/storage resource requirements for an application. However a virtual node in a VDC is limited to a virtual machine (VM), which can only be embedded onto a single physical machine. This condition limits the applicability of the VDC abstraction and the potential of deploying parallel computing. In this paper, we propose a new abstraction based on our Application Centric Network Virtualization (ACNV) approach. Specifically, we model a service application offered by a service provider as a virtual network of service function nodes, which closely matches the service provider s view on the architecture of the application. An infrastructure provider that hosts the application decides how to map the virtual network to the substrate network. Different from the VDC abstraction, each virtual node in our proposed abstraction can be split and mapped onto multiple physical machines, which allows the infrastructure provider to provide auto scaling for the application with variable number of physical machines for exploring the full benefits of parallel computing. We also allow multiple virtual nodes to be mapped and colocated in the same physical machine to minimize resource fragmentation and communication overhead. Extensive simulation results show that the proposed ACNV abstraction outperforms existing VDC-like approaches in achieving optimal resource usage.